{
  "retrievedAt": "2026-10-06T02:02:50.342281+00:00",
  "source": "https://arcprize.org/api/models",
  "requestFilters": {
    "models": [],
    "runners": [],
    "configs": []
  },
  "games": [
    {
      "game": "r11l",
      "source": "https://arcprize.org/tasks/r11l",
      "runs": [
        {
          "session_id": "a45b40f4-2dc4-49bf-9231-ad9231dedf2a",
          "game_id": "r11l",
          "game_version": "495a7899",
          "published_at": "2026-09-25T10:12:56.961339Z",
          "score": 2.2507440476190474,
          "model": "gpt-6-sol",
          "runner": "benchmark_agent",
          "config": "openai-gpt-6-sol-max",
          "verified_publish": true,
          "levels_completed": 1,
          "actions": 197,
          "resets": 3,
          "state": "GAME_OVER",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            47.265625,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            32,
            165,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            22,
            33,
            51,
            26,
            52,
            49
          ]
        },
        {
          "session_id": "d5bbecba-2f2b-43df-a36b-b22de3c717e3",
          "game_id": "r11l",
          "game_version": "495a7899",
          "published_at": "2026-09-25T07:10:33.053437Z",
          "score": 58.82651163728033,
          "model": "gpt-6-sol",
          "runner": "benchmark_agent",
          "config": "openai-gpt-6-sol-xhigh-provider-adapter",
          "verified_publish": true,
          "levels_completed": 5,
          "actions": 433,
          "resets": 5,
          "state": "GAME_OVER",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            115,
            100,
            80.05540166204986,
            66.015625,
            83.22560787934748,
            0
          ],
          "level_actions": [
            9,
            33,
            57,
            32,
            57,
            245
          ],
          "level_baseline_actions": [
            22,
            33,
            51,
            26,
            52,
            49
          ]
        },
        {
          "session_id": "35c3bd1e-2166-49a0-b3db-e08839be35a8",
          "game_id": "r11l",
          "game_version": "495a7899",
          "published_at": "2026-09-25T07:08:53.607137Z",
          "score": 28.579392815225372,
          "model": "gpt-6-sol",
          "runner": "benchmark_agent",
          "config": "openai-gpt-6-sol-high-provider-adapter",
          "verified_publish": true,
          "levels_completed": 4,
          "actions": 506,
          "resets": 14,
          "state": "GAME_OVER",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            57.550535077288934,
            5.179548156956005,
            115,
            46.81440443213297,
            0,
            0
          ],
          "level_actions": [
            29,
            145,
            34,
            38,
            260,
            0
          ],
          "level_baseline_actions": [
            22,
            33,
            51,
            26,
            52,
            49
          ]
        },
        {
          "session_id": "4c760afd-8f27-4978-b95d-82b71d775576",
          "game_id": "r11l",
          "game_version": "495a7899",
          "published_at": "2026-09-25T06:49:48.553052Z",
          "score": 0.7903847410020248,
          "model": "gpt-6-sol",
          "runner": "benchmark_agent",
          "config": "openai-gpt-6-sol-none-provider-adapter",
          "verified_publish": true,
          "levels_completed": 1,
          "actions": 219,
          "resets": 3,
          "state": "GAME_OVER",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            16.59807956104252,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            54,
            165,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            22,
            33,
            51,
            26,
            52,
            49
          ]
        },
        {
          "session_id": "96d0e516-2d3a-416c-b8b0-1f7326f893d4",
          "game_id": "r11l",
          "game_version": "495a7899",
          "published_at": "2026-09-25T06:25:33.558756Z",
          "score": 0,
          "model": "gpt-6-sol",
          "runner": "benchmark_agent",
          "config": "openai-gpt-6-sol-none",
          "verified_publish": true,
          "levels_completed": 0,
          "actions": 110,
          "resets": 1,
          "state": "GAME_OVER",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            110,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            22,
            33,
            51,
            26,
            52,
            49
          ]
        },
        {
          "session_id": "776c46fe-5d73-4d3c-8fff-23057d5d6f6a",
          "game_id": "r11l",
          "game_version": "495a7899",
          "published_at": "2026-09-25T06:22:29.076283Z",
          "score": 0,
          "model": "gpt-6-sol",
          "runner": "benchmark_agent",
          "config": "openai-gpt-6-sol-low-provider-adapter",
          "verified_publish": true,
          "levels_completed": 0,
          "actions": 110,
          "resets": 3,
          "state": "GAME_OVER",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            110,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            22,
            33,
            51,
            26,
            52,
            49
          ]
        },
        {
          "session_id": "966a118b-10b6-4e70-ada6-f9e4f413959a",
          "game_id": "r11l",
          "game_version": "495a7899",
          "published_at": "2026-09-25T06:21:37.418647Z",
          "score": 4.356827797281484,
          "model": "gpt-6-sol",
          "runner": "benchmark_agent",
          "config": "openai-gpt-6-sol-medium-provider-adapter",
          "verified_publish": true,
          "levels_completed": 1,
          "actions": 188,
          "resets": 7,
          "state": "GAME_OVER",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            91.49338374291116,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            23,
            165,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            22,
            33,
            51,
            26,
            52,
            49
          ]
        },
        {
          "session_id": "8a673557-ec0f-4a4d-b2b2-135797fd560b",
          "game_id": "r11l",
          "game_version": "495a7899",
          "published_at": "2026-09-25T06:06:04.010489Z",
          "score": 4.761904761904762,
          "model": "gpt-6-sol",
          "runner": "benchmark_agent",
          "config": "openai-gpt-6-sol-low",
          "verified_publish": true,
          "levels_completed": 1,
          "actions": 177,
          "resets": 3,
          "state": "GAME_OVER",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            115,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            12,
            165,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            22,
            33,
            51,
            26,
            52,
            49
          ]
        },
        {
          "session_id": "64511e1b-d5af-4b92-899e-d8819360ee7f",
          "game_id": "r11l",
          "game_version": "495a7899",
          "published_at": "2026-09-25T05:45:58.513674Z",
          "score": 0.7349368318756073,
          "model": "gpt-6-sol",
          "runner": "benchmark_agent",
          "config": "openai-gpt-6-sol-high",
          "verified_publish": true,
          "levels_completed": 1,
          "actions": 221,
          "resets": 3,
          "state": "GAME_OVER",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            15.433673469387754,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            56,
            165,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            22,
            33,
            51,
            26,
            52,
            49
          ]
        },
        {
          "session_id": "bed13df4-dbab-4e4f-86a7-bfbaad936558",
          "game_id": "r11l",
          "game_version": "495a7899",
          "published_at": "2026-09-25T05:41:05.437022Z",
          "score": 4.761904761904762,
          "model": "gpt-6-sol",
          "runner": "benchmark_agent",
          "config": "openai-gpt-6-sol-xhigh",
          "verified_publish": true,
          "levels_completed": 1,
          "actions": 183,
          "resets": 3,
          "state": "GAME_OVER",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            115,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            18,
            165,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            22,
            33,
            51,
            26,
            52,
            49
          ]
        },
        {
          "session_id": "d46387d0-b068-493b-9ea3-148618629f48",
          "game_id": "r11l",
          "game_version": "495a7899",
          "published_at": "2026-09-25T05:33:10.261202Z",
          "score": 0,
          "model": "gpt-6-sol",
          "runner": "benchmark_agent",
          "config": "openai-gpt-6-sol-medium",
          "verified_publish": true,
          "levels_completed": 0,
          "actions": 110,
          "resets": 1,
          "state": "GAME_OVER",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            110,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            22,
            33,
            51,
            26,
            52,
            49
          ]
        },
        {
          "session_id": "186be5ee-59a1-4f4a-b682-f4333a4ef6fd",
          "game_id": "r11l",
          "game_version": "495a7899",
          "published_at": "2026-09-21T12:18:54.842227Z",
          "score": 4.761904761904762,
          "model": "gemini-3.8-flash",
          "runner": "benchmark_agent",
          "config": "google-gemini-3-8-flash-low-provider-adapter",
          "verified_publish": true,
          "levels_completed": 1,
          "actions": 179,
          "resets": 5,
          "state": "GAME_OVER",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            115,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            14,
            165,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            22,
            33,
            51,
            26,
            52,
            49
          ]
        },
        {
          "session_id": "4b0bd126-d125-4cc7-a3e4-d99352e67cdf",
          "game_id": "r11l",
          "game_version": "495a7899",
          "published_at": "2026-09-20T07:13:53.190565Z",
          "score": 14.285714285714285,
          "model": "gemini-3.8-flash",
          "runner": "benchmark_agent",
          "config": "google-gemini-3-8-flash-medium-provider-adapter",
          "verified_publish": true,
          "levels_completed": 2,
          "actions": 295,
          "resets": 5,
          "state": "GAME_OVER",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            115,
            115,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            14,
            26,
            255,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            22,
            33,
            51,
            26,
            52,
            49
          ]
        },
        {
          "session_id": "6b69c2cf-5cf4-491f-82db-946081ea5eec",
          "game_id": "r11l",
          "game_version": "495a7899",
          "published_at": "2026-09-17T17:30:16.039499Z",
          "score": 89.68239795918367,
          "model": "gemini-3.8-flash",
          "runner": "benchmark_agent",
          "config": "google-gemini-3-8-flash-high-provider-adapter",
          "verified_publish": true,
          "levels_completed": 6,
          "actions": 234,
          "resets": 1,
          "state": "WIN",
          "completed": true,
          "number_of_levels": 0,
          "level_scores": [
            39.51020408163265,
            115,
            82.94005102040816,
            25,
            115,
            115
          ],
          "level_actions": [
            35,
            19,
            56,
            52,
            41,
            31
          ],
          "level_baseline_actions": [
            22,
            33,
            51,
            26,
            52,
            49
          ]
        },
        {
          "session_id": "a8324524-5920-4b39-bebc-b4c2ad4b69d0",
          "game_id": "r11l",
          "game_version": "495a7899",
          "published_at": "2026-09-16T19:37:12.137704Z",
          "score": 0,
          "model": "gpt-6-luna",
          "runner": "benchmark_agent",
          "config": "openai-gpt-6-luna-medium-provider-adapter",
          "verified_publish": true,
          "levels_completed": 0,
          "actions": 110,
          "resets": 2,
          "state": "GAME_OVER",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            110,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            22,
            33,
            51,
            26,
            52,
            49
          ]
        },
        {
          "session_id": "e40b0718-6f3f-41fa-8dcb-ead33e8bcd18",
          "game_id": "r11l",
          "game_version": "495a7899",
          "published_at": "2026-09-16T19:28:18.892491Z",
          "score": 0,
          "model": "gpt-6-luna",
          "runner": "benchmark_agent",
          "config": "openai-gpt-6-luna-low",
          "verified_publish": true,
          "levels_completed": 0,
          "actions": 110,
          "resets": 3,
          "state": "GAME_OVER",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            110,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            22,
            33,
            51,
            26,
            52,
            49
          ]
        },
        {
          "session_id": "bc67c85c-97ad-4988-bf43-d290fc2d5a90",
          "game_id": "r11l",
          "game_version": "495a7899",
          "published_at": "2026-09-16T19:26:16.799567Z",
          "score": 0,
          "model": "gpt-6-luna",
          "runner": "benchmark_agent",
          "config": "openai-gpt-6-luna-none-provider-adapter",
          "verified_publish": true,
          "levels_completed": 0,
          "actions": 110,
          "resets": 1,
          "state": "NOT_FINISHED",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            110,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            22,
            33,
            51,
            26,
            52,
            49
          ]
        },
        {
          "session_id": "88a594aa-e1d4-4125-9b7a-27aedbadf23a",
          "game_id": "r11l",
          "game_version": "495a7899",
          "published_at": "2026-09-16T19:22:35.624843Z",
          "score": 0,
          "model": "gpt-6-luna",
          "runner": "benchmark_agent",
          "config": "openai-gpt-6-luna-none",
          "verified_publish": true,
          "levels_completed": 0,
          "actions": 110,
          "resets": 2,
          "state": "GAME_OVER",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            110,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            22,
            33,
            51,
            26,
            52,
            49
          ]
        },
        {
          "session_id": "ecc49200-1417-4e2c-bf39-35fcb156cee4",
          "game_id": "r11l",
          "game_version": "495a7899",
          "published_at": "2026-09-16T15:09:34.381688Z",
          "score": 0,
          "model": "gpt-6-luna",
          "runner": "benchmark_agent",
          "config": "openai-gpt-6-luna-medium",
          "verified_publish": true,
          "levels_completed": 0,
          "actions": 110,
          "resets": 2,
          "state": "GAME_OVER",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            110,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            22,
            33,
            51,
            26,
            52,
            49
          ]
        },
        {
          "session_id": "3c14dfb4-b48d-4b87-92f7-819b9bd135b8",
          "game_id": "r11l",
          "game_version": "495a7899",
          "published_at": "2026-09-16T09:54:48.87444Z",
          "score": 0.37882345574653264,
          "model": "gpt-6-luna",
          "runner": "benchmark_agent",
          "config": "openai-gpt-6-luna-xhigh-provider-adapter",
          "verified_publish": true,
          "levels_completed": 1,
          "actions": 224,
          "resets": 5,
          "state": "GAME_OVER",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            7.9552925706771855,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            78,
            146,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            22,
            33,
            51,
            26,
            52,
            49
          ]
        },
        {
          "session_id": "246835e2-22a2-4767-ba48-4147c5fc55fc",
          "game_id": "r11l",
          "game_version": "495a7899",
          "published_at": "2026-09-16T09:28:48.635603Z",
          "score": 0.5626860119047619,
          "model": "gpt-6-luna",
          "runner": "benchmark_agent",
          "config": "openai-gpt-6-luna-high-provider-adapter",
          "verified_publish": true,
          "levels_completed": 1,
          "actions": 229,
          "resets": 13,
          "state": "GAME_OVER",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            11.81640625,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            64,
            165,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            22,
            33,
            51,
            26,
            52,
            49
          ]
        },
        {
          "session_id": "cf8d92e6-f3d1-446e-9f0a-57cf79e07d26",
          "game_id": "r11l",
          "game_version": "495a7899",
          "published_at": "2026-09-16T09:18:25.490821Z",
          "score": 0,
          "model": "gpt-6-luna",
          "runner": "benchmark_agent",
          "config": "openai-gpt-6-luna-low-provider-adapter",
          "verified_publish": true,
          "levels_completed": 0,
          "actions": 110,
          "resets": 2,
          "state": "NOT_FINISHED",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            110,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            22,
            33,
            51,
            26,
            52,
            49
          ]
        },
        {
          "session_id": "2d166d81-bdbf-41e8-8c11-ddbe83665012",
          "game_id": "r11l",
          "game_version": "495a7899",
          "published_at": "2026-09-16T09:06:56.554415Z",
          "score": 0,
          "model": "gpt-6-luna",
          "runner": "benchmark_agent",
          "config": "openai-gpt-6-luna-max",
          "verified_publish": true,
          "levels_completed": 0,
          "actions": 110,
          "resets": 4,
          "state": "GAME_OVER",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            110,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            22,
            33,
            51,
            26,
            52,
            49
          ]
        },
        {
          "session_id": "552802c7-ab2f-4eed-a0a4-daee0f98840c",
          "game_id": "r11l",
          "game_version": "495a7899",
          "published_at": "2026-09-16T08:58:28.948409Z",
          "score": 0,
          "model": "gpt-6-luna",
          "runner": "benchmark_agent",
          "config": "openai-gpt-6-luna-high",
          "verified_publish": true,
          "levels_completed": 0,
          "actions": 110,
          "resets": 2,
          "state": "GAME_OVER",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            110,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            22,
            33,
            51,
            26,
            52,
            49
          ]
        },
        {
          "session_id": "a78cc9f7-1a81-4f00-bcef-2403b0b3d40c",
          "game_id": "r11l",
          "game_version": "495a7899",
          "published_at": "2026-09-16T08:01:21.907718Z",
          "score": 0,
          "model": "gpt-6-luna",
          "runner": "benchmark_agent",
          "config": "openai-gpt-6-luna-xhigh",
          "verified_publish": true,
          "levels_completed": 0,
          "actions": 110,
          "resets": 2,
          "state": "GAME_OVER",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            110,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            22,
            33,
            51,
            26,
            52,
            49
          ]
        },
        {
          "session_id": "cccf10e5-ff8a-4a62-bc95-3cb22c0b81f5",
          "game_id": "r11l",
          "game_version": "495a7899",
          "published_at": "2026-09-16T07:11:59.283446Z",
          "score": 1.0003306878306877,
          "model": "gpt-6-1-sol",
          "runner": "benchmark_agent",
          "config": "openai-gpt-6-1-sol-low",
          "verified_publish": true,
          "levels_completed": 1,
          "actions": 213,
          "resets": 3,
          "state": "GAME_OVER",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            21.006944444444443,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            48,
            165,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            22,
            33,
            51,
            26,
            52,
            49
          ]
        },
        {
          "session_id": "ccc817cf-cb57-4622-aca3-fff6c872cb93",
          "game_id": "r11l",
          "game_version": "495a7899",
          "published_at": "2026-09-16T06:20:36.569314Z",
          "score": 5.6751193310283705,
          "model": "gpt-6-1-sol",
          "runner": "benchmark_agent",
          "config": "openai-gpt-6-1-sol-medium",
          "verified_publish": true,
          "levels_completed": 2,
          "actions": 428,
          "resets": 13,
          "state": "GAME_OVER",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            109.75056689342404,
            4.713469529085873,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            21,
            152,
            255,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            22,
            33,
            51,
            26,
            52,
            49
          ]
        },
        {
          "session_id": "063d5304-1ee7-4527-8963-f01ce0648731",
          "game_id": "r11l",
          "game_version": "495a7899",
          "published_at": "2026-09-16T05:29:28.395613Z",
          "score": 59.90579607194319,
          "model": "gpt-6-1-sol",
          "runner": "benchmark_agent",
          "config": "openai-gpt-6-1-sol-high",
          "verified_publish": true,
          "levels_completed": 6,
          "actions": 398,
          "resets": 9,
          "state": "WIN",
          "completed": true,
          "number_of_levels": 0,
          "level_scores": [
            115,
            115,
            115,
            115,
            7.900657414170928,
            11.41973840665874
          ],
          "level_actions": [
            6,
            20,
            21,
            21,
            185,
            145
          ],
          "level_baseline_actions": [
            22,
            33,
            51,
            26,
            52,
            49
          ]
        },
        {
          "session_id": "2ed50725-ff52-42a5-854b-14ff4bdb3891",
          "game_id": "r11l",
          "game_version": "495a7899",
          "published_at": "2026-09-16T04:54:16.411052Z",
          "score": 100,
          "model": "gpt-6-1-sol",
          "runner": "benchmark_agent",
          "config": "openai-gpt-6-1-sol-high-provider-adapter",
          "verified_publish": true,
          "levels_completed": 6,
          "actions": 101,
          "resets": 0,
          "state": "WIN",
          "completed": true,
          "number_of_levels": 0,
          "level_scores": [
            115,
            115,
            115,
            115,
            115,
            115
          ],
          "level_actions": [
            8,
            11,
            12,
            14,
            38,
            18
          ],
          "level_baseline_actions": [
            22,
            33,
            51,
            26,
            52,
            49
          ]
        },
        {
          "session_id": "545c0196-a9a8-43b1-aa53-3f825f9f98a4",
          "game_id": "r11l",
          "game_version": "495a7899",
          "published_at": "2026-09-16T04:51:53.106229Z",
          "score": 0.25537528030602824,
          "model": "gpt-6-luna",
          "runner": "benchmark_agent",
          "config": "openai-gpt-6-luna-max-provider-adapter",
          "verified_publish": true,
          "levels_completed": 1,
          "actions": 260,
          "resets": 13,
          "state": "GAME_OVER",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            5.362880886426593,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            95,
            165,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            22,
            33,
            51,
            26,
            52,
            49
          ]
        },
        {
          "session_id": "7d4a9d4c-e1d4-441c-8444-53a9f8399766",
          "game_id": "r11l",
          "game_version": "495a7899",
          "published_at": "2026-09-16T04:35:24.799709Z",
          "score": 100,
          "model": "gpt-6-1-sol",
          "runner": "benchmark_agent",
          "config": "openai-gpt-6-1-sol-max",
          "verified_publish": true,
          "levels_completed": 6,
          "actions": 98,
          "resets": 0,
          "state": "WIN",
          "completed": true,
          "number_of_levels": 0,
          "level_scores": [
            115,
            115,
            115,
            115,
            115,
            115
          ],
          "level_actions": [
            6,
            14,
            18,
            15,
            22,
            23
          ],
          "level_baseline_actions": [
            22,
            33,
            51,
            26,
            52,
            49
          ]
        },
        {
          "session_id": "90b7141f-0ff2-4a36-8cf4-2fa32ec23dee",
          "game_id": "r11l",
          "game_version": "495a7899",
          "published_at": "2026-09-16T03:37:32.297542Z",
          "score": 100,
          "model": "gpt-6-1-sol",
          "runner": "benchmark_agent",
          "config": "openai-gpt-6-1-sol-max-provider-adapter",
          "verified_publish": true,
          "levels_completed": 6,
          "actions": 98,
          "resets": 0,
          "state": "WIN",
          "completed": true,
          "number_of_levels": 0,
          "level_scores": [
            115,
            115,
            115,
            115,
            115,
            115
          ],
          "level_actions": [
            4,
            14,
            27,
            17,
            21,
            15
          ],
          "level_baseline_actions": [
            22,
            33,
            51,
            26,
            52,
            49
          ]
        },
        {
          "session_id": "8d26f8f7-b181-4c6b-b63d-82546928f3ed",
          "game_id": "r11l",
          "game_version": "495a7899",
          "published_at": "2026-09-16T03:18:33.668035Z",
          "score": 100,
          "model": "gpt-6-1-sol",
          "runner": "benchmark_agent",
          "config": "openai-gpt-6-1-sol-xhigh",
          "verified_publish": true,
          "levels_completed": 6,
          "actions": 119,
          "resets": 2,
          "state": "WIN",
          "completed": true,
          "number_of_levels": 0,
          "level_scores": [
            115,
            115,
            115,
            115,
            112.61974177426073,
            115
          ],
          "level_actions": [
            3,
            8,
            22,
            14,
            49,
            23
          ],
          "level_baseline_actions": [
            22,
            33,
            51,
            26,
            52,
            49
          ]
        },
        {
          "session_id": "34034051-d97e-4709-b77c-d6d2039a537c",
          "game_id": "r11l",
          "game_version": "495a7899",
          "published_at": "2026-09-16T03:00:45.220982Z",
          "score": 100,
          "model": "gpt-6-1-sol",
          "runner": "benchmark_agent",
          "config": "openai-gpt-6-1-sol-xhigh-provider-adapter",
          "verified_publish": true,
          "levels_completed": 6,
          "actions": 113,
          "resets": 1,
          "state": "WIN",
          "completed": true,
          "number_of_levels": 0,
          "level_scores": [
            115,
            115,
            115,
            115,
            115,
            115
          ],
          "level_actions": [
            4,
            13,
            22,
            18,
            21,
            35
          ],
          "level_baseline_actions": [
            22,
            33,
            51,
            26,
            52,
            49
          ]
        },
        {
          "session_id": "d37d2c61-3128-4833-be85-b406385fd162",
          "game_id": "r11l",
          "game_version": "495a7899",
          "published_at": "2026-09-12T09:19:31.820435Z",
          "score": 4.761904761904762,
          "model": "gemini-3.8-flash",
          "runner": "benchmark_agent",
          "config": "google-gemini-3-8-flash-low",
          "verified_publish": true,
          "levels_completed": 1,
          "actions": 169,
          "resets": 5,
          "state": "GAME_OVER",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            115,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            4,
            165,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            22,
            33,
            51,
            26,
            52,
            49
          ]
        },
        {
          "session_id": "202f3ba7-fdbf-4b1f-af3c-bb857fe8f11a",
          "game_id": "r11l",
          "game_version": "495a7899",
          "published_at": "2026-09-10T23:52:01.27336Z",
          "score": 0.8204919561274137,
          "model": "gemini-3.8-flash",
          "runner": "benchmark_agent",
          "config": "google-gemini-3-8-flash-medium",
          "verified_publish": true,
          "levels_completed": 1,
          "actions": 218,
          "resets": 4,
          "state": "GAME_OVER",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            17.230331078675686,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            53,
            165,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            22,
            33,
            51,
            26,
            52,
            49
          ]
        },
        {
          "session_id": "8feacd51-98e0-4536-8ced-3fa7833f5499",
          "game_id": "r11l",
          "game_version": "495a7899",
          "published_at": "2026-09-10T22:29:20.827306Z",
          "score": 22.613596342481475,
          "model": "gemini-3.8-flash",
          "runner": "benchmark_agent",
          "config": "google-gemini-3-8-flash-high",
          "verified_publish": true,
          "levels_completed": 3,
          "actions": 232,
          "resets": 6,
          "state": "GAME_OVER",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            35.35427319211104,
            47.265625,
            115,
            0,
            0,
            0
          ],
          "level_actions": [
            37,
            48,
            17,
            130,
            0,
            0
          ],
          "level_baseline_actions": [
            22,
            33,
            51,
            26,
            52,
            49
          ]
        },
        {
          "session_id": "02035a34-e3ba-475b-9c4a-5afad226af02",
          "game_id": "r11l",
          "game_version": "495a7899",
          "published_at": "2026-08-11T10:38:21.768134Z",
          "score": 0,
          "model": "grok-4.6",
          "runner": "benchmark_agent",
          "config": "xai-grok-4-6-xhigh",
          "verified_publish": true,
          "levels_completed": 0,
          "actions": 110,
          "resets": 1,
          "state": "GAME_OVER",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            110,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            22,
            33,
            51,
            26,
            52,
            49
          ]
        },
        {
          "session_id": "f2270566-2d3d-42d4-b7e7-02acceab2a7f",
          "game_id": "r11l",
          "game_version": "495a7899",
          "published_at": "2026-07-21T10:14:18.49026Z",
          "score": 100,
          "model": "claude-opus-5",
          "runner": "benchmark_agent",
          "config": "anthropic-claude-opus-5-high",
          "verified_publish": true,
          "levels_completed": 6,
          "actions": 117,
          "resets": 1,
          "state": "WIN",
          "completed": true,
          "number_of_levels": 0,
          "level_scores": [
            115,
            113.3194588969823,
            115,
            115,
            115,
            115
          ],
          "level_actions": [
            3,
            31,
            18,
            18,
            27,
            20
          ],
          "level_baseline_actions": [
            22,
            33,
            51,
            26,
            52,
            49
          ]
        },
        {
          "session_id": "2c9e6a5e-d011-48e9-b216-356fc03da54d",
          "game_id": "r11l",
          "game_version": "495a7899",
          "published_at": "2026-06-24T10:58:17.002395Z",
          "score": 0,
          "model": "gpt-5-6-luna",
          "runner": "benchmark_agent",
          "config": "openai-gpt-5-6-luna-low",
          "verified_publish": true,
          "levels_completed": 0,
          "actions": 110,
          "resets": 1,
          "state": "GAME_OVER",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            110,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            22,
            33,
            51,
            26,
            52,
            49
          ]
        },
        {
          "session_id": "e5bd5047-a84b-4615-a991-4d536f43277a",
          "game_id": "r11l",
          "game_version": "495a7899",
          "published_at": "2026-06-24T10:43:56.879204Z",
          "score": 0,
          "model": "gpt-5-6-luna",
          "runner": "benchmark_agent",
          "config": "openai-gpt-5-6-luna-medium",
          "verified_publish": true,
          "levels_completed": 0,
          "actions": 110,
          "resets": 3,
          "state": "GAME_OVER",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            110,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            22,
            33,
            51,
            26,
            52,
            49
          ]
        },
        {
          "session_id": "5d4c4dbd-fb65-4120-829e-86c90d2f85d9",
          "game_id": "r11l",
          "game_version": "495a7899",
          "published_at": "2026-06-24T04:47:13.093987Z",
          "score": 0,
          "model": "gpt-5-6-luna",
          "runner": "benchmark_agent",
          "config": "openai-gpt-5-6-luna-high",
          "verified_publish": true,
          "levels_completed": 0,
          "actions": 110,
          "resets": 5,
          "state": "NOT_FINISHED",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            110,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            22,
            33,
            51,
            26,
            52,
            49
          ]
        },
        {
          "session_id": "eccb6660-37cd-4164-b5f1-3e3de6d74a7e",
          "game_id": "r11l",
          "game_version": "495a7899",
          "published_at": "2026-06-24T03:19:10.691833Z",
          "score": 0,
          "model": "gpt-5-6-luna",
          "runner": "benchmark_agent",
          "config": "openai-gpt-5-6-luna-xhigh",
          "verified_publish": true,
          "levels_completed": 0,
          "actions": 110,
          "resets": 8,
          "state": "NOT_FINISHED",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            110,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            22,
            33,
            51,
            26,
            52,
            49
          ]
        },
        {
          "session_id": "4fd2ae50-0602-4176-9517-740c92c35293",
          "game_id": "r11l",
          "game_version": "495a7899",
          "published_at": "2026-06-23T05:27:21.147911Z",
          "score": 0,
          "model": "gpt-5-6-luna",
          "runner": "benchmark_agent",
          "config": "openai-gpt-5-6-luna-max",
          "verified_publish": true,
          "levels_completed": 0,
          "actions": 110,
          "resets": 9,
          "state": "NOT_FINISHED",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            110,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            22,
            33,
            51,
            26,
            52,
            49
          ]
        },
        {
          "session_id": "afba813c-8f0e-42dd-927e-e845e3f2840d",
          "game_id": "r11l",
          "game_version": "495a7899",
          "published_at": "2026-06-20T12:05:09.291815Z",
          "score": 0,
          "model": "gpt-5-6-terra",
          "runner": "benchmark_agent",
          "config": "openai-gpt-5-6-terra-low",
          "verified_publish": true,
          "levels_completed": 0,
          "actions": 110,
          "resets": 1,
          "state": "GAME_OVER",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            110,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            22,
            33,
            51,
            26,
            52,
            49
          ]
        },
        {
          "session_id": "f80f5a09-fa50-4f9b-a301-e68bc91c2556",
          "game_id": "r11l",
          "game_version": "495a7899",
          "published_at": "2026-06-20T09:32:55.776401Z",
          "score": 0,
          "model": "gpt-5-6-terra",
          "runner": "benchmark_agent",
          "config": "openai-gpt-5-6-terra-medium",
          "verified_publish": true,
          "levels_completed": 0,
          "actions": 110,
          "resets": 1,
          "state": "GAME_OVER",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            110,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            22,
            33,
            51,
            26,
            52,
            49
          ]
        },
        {
          "session_id": "ccb9d6b2-cfc4-454c-b228-35f0fe62892d",
          "game_id": "r11l",
          "game_version": "495a7899",
          "published_at": "2026-06-20T09:19:29.753362Z",
          "score": 0,
          "model": "gpt-5-6-terra",
          "runner": "benchmark_agent",
          "config": "openai-gpt-5-6-terra-high",
          "verified_publish": true,
          "levels_completed": 0,
          "actions": 110,
          "resets": 1,
          "state": "GAME_OVER",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            110,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            22,
            33,
            51,
            26,
            52,
            49
          ]
        },
        {
          "session_id": "71997ec5-8842-4e9c-8906-e4f279bdf0cb",
          "game_id": "r11l",
          "game_version": "495a7899",
          "published_at": "2026-06-20T05:44:31.168206Z",
          "score": 4.356827797281484,
          "model": "gpt-5-6-terra",
          "runner": "benchmark_agent",
          "config": "openai-gpt-5-6-terra-xhigh",
          "verified_publish": true,
          "levels_completed": 1,
          "actions": 188,
          "resets": 3,
          "state": "GAME_OVER",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            91.49338374291116,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            23,
            165,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            22,
            33,
            51,
            26,
            52,
            49
          ]
        },
        {
          "session_id": "862d31d3-2bf3-44c7-8333-ed14abe15a69",
          "game_id": "r11l",
          "game_version": "495a7899",
          "published_at": "2026-06-18T23:38:32.1191Z",
          "score": 0,
          "model": "gpt-5-6-terra",
          "runner": "benchmark_agent",
          "config": "openai-gpt-5-6-terra-max",
          "verified_publish": true,
          "levels_completed": 0,
          "actions": 110,
          "resets": 15,
          "state": "NOT_FINISHED",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            110,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            22,
            33,
            51,
            26,
            52,
            49
          ]
        },
        {
          "session_id": "47fa0f6e-3b83-4830-914b-0a49592b8ca8",
          "game_id": "r11l",
          "game_version": "495a7899",
          "published_at": "2026-06-15T19:47:22.961573Z",
          "score": 2.5608465608465605,
          "model": "gpt-5-6-sol",
          "runner": "benchmark_agent",
          "config": "openai-gpt-5-6-sol-xhigh",
          "verified_publish": true,
          "levels_completed": 1,
          "actions": 195,
          "resets": 9,
          "state": "GAME_OVER",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            53.77777777777777,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            30,
            165,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            22,
            33,
            51,
            26,
            52,
            49
          ]
        },
        {
          "session_id": "ff216a27-e10e-4e88-90e9-c08d99017976",
          "game_id": "r11l",
          "game_version": "495a7899",
          "published_at": "2026-06-13T09:59:19.360236Z",
          "score": 14.285714285714285,
          "model": "gpt-5-6-sol",
          "runner": "benchmark_agent",
          "config": "openai-gpt-5-6-sol-max",
          "verified_publish": true,
          "levels_completed": 2,
          "actions": 286,
          "resets": 11,
          "state": "GAME_OVER",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            115,
            115,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            18,
            13,
            255,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            22,
            33,
            51,
            26,
            52,
            49
          ]
        },
        {
          "session_id": "09a2e5b9-cd6a-4aac-aba5-dde44efc7934",
          "game_id": "r11l",
          "game_version": "495a7899",
          "published_at": "2026-06-12T07:02:27.323651Z",
          "score": 0.19398719844810244,
          "model": "gpt-5-6-sol",
          "runner": "benchmark_agent",
          "config": "openai-gpt-5-6-sol-high",
          "verified_publish": true,
          "levels_completed": 1,
          "actions": 274,
          "resets": 7,
          "state": "GAME_OVER",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            4.073731167410151,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            109,
            165,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            22,
            33,
            51,
            26,
            52,
            49
          ]
        },
        {
          "session_id": "26ba4ebf-9ac2-45ce-a88e-a9f0c40f8edd",
          "game_id": "r11l",
          "game_version": "495a7899",
          "published_at": "2026-06-10T09:07:31.348479Z",
          "score": 0,
          "model": "gpt-5-6-sol",
          "runner": "benchmark_agent",
          "config": "openai-gpt-5-6-sol-medium",
          "verified_publish": true,
          "levels_completed": 0,
          "actions": 110,
          "resets": 1,
          "state": "GAME_OVER",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            110,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            22,
            33,
            51,
            26,
            52,
            49
          ]
        },
        {
          "session_id": "2c24425d-c7ac-4956-9dc5-60c30ff6bc8a",
          "game_id": "r11l",
          "game_version": "495a7899",
          "published_at": "2026-06-09T22:50:23.009176Z",
          "score": 0.2130881938574246,
          "model": "gpt-5-6-sol",
          "runner": "benchmark_agent",
          "config": "openai-gpt-5-6-sol-low",
          "verified_publish": true,
          "levels_completed": 1,
          "actions": 269,
          "resets": 3,
          "state": "GAME_OVER",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            4.474852071005917,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            104,
            165,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            22,
            33,
            51,
            26,
            52,
            49
          ]
        },
        {
          "session_id": "57231fb9-f620-4677-8363-1b3b82e1308e",
          "game_id": "r11l",
          "game_version": "495a7899",
          "published_at": "2026-05-29T17:06:50.530876Z",
          "score": 0,
          "model": "claude-opus-4-8",
          "runner": "benchmark_agent",
          "config": "anthropic-opus-4-8-high",
          "verified_publish": true,
          "levels_completed": 0,
          "actions": 110,
          "resets": 3,
          "state": "GAME_OVER",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            110,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            22,
            33,
            51,
            26,
            52,
            49
          ]
        },
        {
          "session_id": "c13178e6-964c-464b-8846-2e9e2a621272",
          "game_id": "r11l",
          "game_version": "495a7899",
          "published_at": "2026-05-01T05:07:31.956942Z",
          "score": 0,
          "model": "claude-opus-4-7",
          "runner": "benchmark_agent",
          "config": "anthropic-opus-4-7-high",
          "verified_publish": true,
          "levels_completed": 0,
          "actions": 110,
          "resets": 3,
          "state": "GAME_OVER",
          "completed": false,
          "number_of_levels": 6,
          "level_scores": [
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            110,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            22,
            33,
            51,
            26,
            52,
            49
          ]
        },
        {
          "session_id": "a0023983-f942-4ead-a282-d97020cec97d",
          "game_id": "r11l",
          "game_version": "495a7899",
          "published_at": "2026-04-30T23:21:04.982248Z",
          "score": 0,
          "model": "gpt-5.5-2026-04-23",
          "runner": "benchmark_agent",
          "config": "openai-gpt-5-5-2026-04-23-high",
          "verified_publish": true,
          "levels_completed": 0,
          "actions": 110,
          "resets": 1,
          "state": "GAME_OVER",
          "completed": false,
          "number_of_levels": 6,
          "level_scores": [
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            110,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            22,
            33,
            51,
            26,
            52,
            49
          ]
        },
        {
          "session_id": "a34c6ef0-ff91-4de4-85e0-0b957b27e7d8",
          "game_id": "r11l",
          "game_version": "495a7899",
          "published_at": "2026-04-14T17:44:22.989088Z",
          "score": 0,
          "model": "grok-4.20-beta-0309-reasoning",
          "runner": "benchmark_agent",
          "config": "xai-grok-4-20-beta-0309-reasoning",
          "verified_publish": true,
          "levels_completed": 0,
          "actions": 55,
          "resets": 0,
          "state": "NOT_FINISHED",
          "completed": false,
          "number_of_levels": 6,
          "level_scores": [
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            55,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            22,
            33,
            51,
            26,
            52,
            49
          ]
        },
        {
          "session_id": "8d68d10a-fe10-4c29-a9f3-443a036974c0",
          "game_id": "r11l",
          "game_version": "495a7899",
          "published_at": "2026-04-14T17:44:22.419609Z",
          "score": 0,
          "model": "gpt-5.4-2026-03-05",
          "runner": "benchmark_agent",
          "config": "openai-gpt-5-4-2026-03-05-high",
          "verified_publish": true,
          "levels_completed": 0,
          "actions": 110,
          "resets": 1,
          "state": "GAME_OVER",
          "completed": false,
          "number_of_levels": 6,
          "level_scores": [
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            110,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            22,
            33,
            51,
            26,
            52,
            49
          ]
        },
        {
          "session_id": "f40febfa-2d23-4b34-a86c-dcee54961c99",
          "game_id": "r11l",
          "game_version": "495a7899",
          "published_at": "2026-04-14T17:44:21.592873Z",
          "score": 4.761904761904762,
          "model": "gemini-3.1-pro-preview",
          "runner": "benchmark_agent",
          "config": "google-gemini-3-1-pro-preview",
          "verified_publish": true,
          "levels_completed": 1,
          "actions": 29,
          "resets": 2,
          "state": "GAME_OVER",
          "completed": false,
          "number_of_levels": 6,
          "level_scores": [
            115,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            4,
            25,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            22,
            33,
            51,
            26,
            52,
            49
          ]
        },
        {
          "session_id": "059a4446-8abd-4a01-b97b-83d758446695",
          "game_id": "r11l",
          "game_version": "495a7899",
          "published_at": "2026-04-14T17:44:21.137123Z",
          "score": 0,
          "model": "claude-opus-4-6",
          "runner": "benchmark_agent",
          "config": "anthropic-opus-4-6-max-effort",
          "verified_publish": true,
          "levels_completed": 0,
          "actions": 110,
          "resets": 12,
          "state": "NOT_FINISHED",
          "completed": false,
          "number_of_levels": 6,
          "level_scores": [
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            110,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            22,
            33,
            51,
            26,
            52,
            49
          ]
        }
      ]
    },
    {
      "game": "su15",
      "source": "https://arcprize.org/tasks/su15",
      "runs": [
        {
          "session_id": "aa06e062-2621-4a45-8a48-7e1d5a91e2f4",
          "game_id": "su15",
          "game_version": "1944f8ab",
          "published_at": "2026-09-25T06:49:33.394001Z",
          "score": 0.8299039780521263,
          "model": "gpt-6-sol",
          "runner": "benchmark_agent",
          "config": "openai-gpt-6-sol-none-provider-adapter",
          "verified_publish": true,
          "levels_completed": 1,
          "actions": 246,
          "resets": 7,
          "state": "NOT_FINISHED",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            37.345679012345684,
            0,
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            36,
            210,
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            22,
            42,
            26,
            115,
            36,
            31,
            8,
            40,
            41
          ]
        },
        {
          "session_id": "8015dd39-b877-4902-a585-ccda4e867f9e",
          "game_id": "su15",
          "game_version": "1944f8ab",
          "published_at": "2026-09-25T06:33:54.629044Z",
          "score": 0.5555555555555556,
          "model": "gpt-6-sol",
          "runner": "benchmark_agent",
          "config": "openai-gpt-6-sol-low-provider-adapter",
          "verified_publish": true,
          "levels_completed": 1,
          "actions": 254,
          "resets": 6,
          "state": "NOT_FINISHED",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            25,
            0,
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            44,
            210,
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            22,
            42,
            26,
            115,
            36,
            31,
            8,
            40,
            41
          ]
        },
        {
          "session_id": "9eb23499-4d66-4678-b88d-3e955ad31fe4",
          "game_id": "su15",
          "game_version": "1944f8ab",
          "published_at": "2026-09-25T06:15:14.338118Z",
          "score": 0,
          "model": "gpt-6-sol",
          "runner": "benchmark_agent",
          "config": "openai-gpt-6-sol-none",
          "verified_publish": true,
          "levels_completed": 0,
          "actions": 110,
          "resets": 2,
          "state": "GAME_OVER",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            0,
            0,
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            110,
            0,
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            22,
            42,
            26,
            115,
            36,
            31,
            8,
            40,
            41
          ]
        },
        {
          "session_id": "3fad3468-715c-4cf8-ab86-5a6081c4f33b",
          "game_id": "su15",
          "game_version": "1944f8ab",
          "published_at": "2026-09-25T06:12:56.080609Z",
          "score": 0.25456936226167,
          "model": "gpt-6-sol",
          "runner": "benchmark_agent",
          "config": "openai-gpt-6-sol-medium-provider-adapter",
          "verified_publish": true,
          "levels_completed": 1,
          "actions": 275,
          "resets": 7,
          "state": "GAME_OVER",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            11.45562130177515,
            0,
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            65,
            210,
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            22,
            42,
            26,
            115,
            36,
            31,
            8,
            40,
            41
          ]
        },
        {
          "session_id": "1f26cab9-f7cf-44b2-8535-b0a74cdf3faf",
          "game_id": "su15",
          "game_version": "1944f8ab",
          "published_at": "2026-09-25T05:50:53.029044Z",
          "score": 0,
          "model": "gpt-6-sol",
          "runner": "benchmark_agent",
          "config": "openai-gpt-6-sol-low",
          "verified_publish": true,
          "levels_completed": 0,
          "actions": 110,
          "resets": 1,
          "state": "GAME_OVER",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            0,
            0,
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            110,
            0,
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            22,
            42,
            26,
            115,
            36,
            31,
            8,
            40,
            41
          ]
        },
        {
          "session_id": "1303da84-e09a-4867-bf54-fa94b5d32f10",
          "game_id": "su15",
          "game_version": "1944f8ab",
          "published_at": "2026-09-25T05:08:04.910488Z",
          "score": 0,
          "model": "gpt-6-sol",
          "runner": "benchmark_agent",
          "config": "openai-gpt-6-sol-medium",
          "verified_publish": true,
          "levels_completed": 0,
          "actions": 110,
          "resets": 2,
          "state": "GAME_OVER",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            0,
            0,
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            110,
            0,
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            22,
            42,
            26,
            115,
            36,
            31,
            8,
            40,
            41
          ]
        },
        {
          "session_id": "6cfd5166-9753-4041-8587-e3efaba3a4e3",
          "game_id": "su15",
          "game_version": "1944f8ab",
          "published_at": "2026-09-25T05:02:56.768554Z",
          "score": 2.2222222222222223,
          "model": "gpt-6-sol",
          "runner": "benchmark_agent",
          "config": "openai-gpt-6-sol-xhigh-provider-adapter",
          "verified_publish": true,
          "levels_completed": 1,
          "actions": 226,
          "resets": 4,
          "state": "GAME_OVER",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            115,
            0,
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            16,
            210,
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            22,
            42,
            26,
            115,
            36,
            31,
            8,
            40,
            41
          ]
        },
        {
          "session_id": "f4f40b9a-2611-4f33-a799-d48f30dda8b2",
          "game_id": "su15",
          "game_version": "1944f8ab",
          "published_at": "2026-09-25T05:02:08.766574Z",
          "score": 0,
          "model": "gpt-6-sol",
          "runner": "benchmark_agent",
          "config": "openai-gpt-6-sol-max",
          "verified_publish": true,
          "levels_completed": 0,
          "actions": 110,
          "resets": 2,
          "state": "NOT_FINISHED",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            0,
            0,
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            110,
            0,
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            22,
            42,
            26,
            115,
            36,
            31,
            8,
            40,
            41
          ]
        },
        {
          "session_id": "f91aa396-f4c8-4675-98ab-70a1eae5fea6",
          "game_id": "su15",
          "game_version": "1944f8ab",
          "published_at": "2026-09-25T04:53:47.892301Z",
          "score": 0,
          "model": "gpt-6-sol",
          "runner": "benchmark_agent",
          "config": "openai-gpt-6-sol-high-provider-adapter",
          "verified_publish": true,
          "levels_completed": 0,
          "actions": 110,
          "resets": 3,
          "state": "NOT_FINISHED",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            0,
            0,
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            110,
            0,
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            22,
            42,
            26,
            115,
            36,
            31,
            8,
            40,
            41
          ]
        },
        {
          "session_id": "b0cb1007-f652-48f4-a21e-93b63934cf63",
          "game_id": "su15",
          "game_version": "1944f8ab",
          "published_at": "2026-09-25T04:25:24.441472Z",
          "score": 0,
          "model": "gpt-6-sol",
          "runner": "benchmark_agent",
          "config": "openai-gpt-6-sol-high",
          "verified_publish": true,
          "levels_completed": 0,
          "actions": 110,
          "resets": 2,
          "state": "GAME_OVER",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            0,
            0,
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            110,
            0,
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            22,
            42,
            26,
            115,
            36,
            31,
            8,
            40,
            41
          ]
        },
        {
          "session_id": "17006008-109d-46c9-8ac3-3e86e1156fc8",
          "game_id": "su15",
          "game_version": "1944f8ab",
          "published_at": "2026-09-25T04:12:45.774709Z",
          "score": 0,
          "model": "gpt-6-sol",
          "runner": "benchmark_agent",
          "config": "openai-gpt-6-sol-xhigh",
          "verified_publish": true,
          "levels_completed": 0,
          "actions": 110,
          "resets": 2,
          "state": "GAME_OVER",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            0,
            0,
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            110,
            0,
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            22,
            42,
            26,
            115,
            36,
            31,
            8,
            40,
            41
          ]
        },
        {
          "session_id": "e18cc8b0-fdd0-4856-a4a8-aad4f8055849",
          "game_id": "su15",
          "game_version": "1944f8ab",
          "published_at": "2026-09-21T17:30:06.219817Z",
          "score": 15.803555555555555,
          "model": "gemini-3.8-flash",
          "runner": "benchmark_agent",
          "config": "google-gemini-3-8-flash-low-provider-adapter",
          "verified_publish": true,
          "levels_completed": 4,
          "actions": 729,
          "resets": 76,
          "state": "GAME_OVER",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            115,
            115,
            115,
            5.29,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            9,
            23,
            17,
            500,
            180,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            22,
            42,
            26,
            115,
            36,
            31,
            8,
            40,
            41
          ]
        },
        {
          "session_id": "aa2d0ddd-d1e6-4fc8-8bd8-3fed3a88141c",
          "game_id": "su15",
          "game_version": "1944f8ab",
          "published_at": "2026-09-20T11:44:47.336452Z",
          "score": 29.42548299441608,
          "model": "gemini-3.8-flash",
          "runner": "benchmark_agent",
          "config": "google-gemini-3-8-flash-medium-provider-adapter",
          "verified_publish": true,
          "levels_completed": 5,
          "actions": 340,
          "resets": 18,
          "state": "GAME_OVER",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            115,
            115,
            115,
            115,
            34.82934694974469,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            8,
            18,
            18,
            80,
            61,
            155,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            22,
            42,
            26,
            115,
            36,
            31,
            8,
            40,
            41
          ]
        },
        {
          "session_id": "f46e1955-181a-42e6-8185-64cfd33485cc",
          "game_id": "su15",
          "game_version": "1944f8ab",
          "published_at": "2026-09-18T00:02:12.907306Z",
          "score": 89.5244177560935,
          "model": "gemini-3.8-flash",
          "runner": "benchmark_agent",
          "config": "google-gemini-3-8-flash-high-provider-adapter",
          "verified_publish": true,
          "levels_completed": 9,
          "actions": 232,
          "resets": 11,
          "state": "WIN",
          "completed": true,
          "number_of_levels": 0,
          "level_scores": [
            115,
            115,
            115,
            115,
            24.319759804841432,
            115,
            16.000000000000004,
            115,
            115
          ],
          "level_actions": [
            11,
            16,
            21,
            16,
            73,
            14,
            20,
            27,
            34
          ],
          "level_baseline_actions": [
            22,
            42,
            26,
            115,
            36,
            31,
            8,
            40,
            41
          ]
        },
        {
          "session_id": "e7361b94-3172-439d-9bd9-dbc4c328a324",
          "game_id": "su15",
          "game_version": "1944f8ab",
          "published_at": "2026-09-16T19:36:21.533709Z",
          "score": 0,
          "model": "gpt-6-luna",
          "runner": "benchmark_agent",
          "config": "openai-gpt-6-luna-medium-provider-adapter",
          "verified_publish": true,
          "levels_completed": 0,
          "actions": 110,
          "resets": 3,
          "state": "GAME_OVER",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            0,
            0,
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            110,
            0,
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            22,
            42,
            26,
            115,
            36,
            31,
            8,
            40,
            41
          ]
        },
        {
          "session_id": "6b0dc06b-8488-4f62-ac88-abdd86030adf",
          "game_id": "su15",
          "game_version": "1944f8ab",
          "published_at": "2026-09-16T19:29:28.515043Z",
          "score": 0,
          "model": "gpt-6-luna",
          "runner": "benchmark_agent",
          "config": "openai-gpt-6-luna-none-provider-adapter",
          "verified_publish": true,
          "levels_completed": 0,
          "actions": 110,
          "resets": 2,
          "state": "GAME_OVER",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            0,
            0,
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            110,
            0,
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            22,
            42,
            26,
            115,
            36,
            31,
            8,
            40,
            41
          ]
        },
        {
          "session_id": "8a1303b7-1491-4b51-9c10-8ee8f2bad26d",
          "game_id": "su15",
          "game_version": "1944f8ab",
          "published_at": "2026-09-16T19:29:10.476853Z",
          "score": 0,
          "model": "gpt-6-luna",
          "runner": "benchmark_agent",
          "config": "openai-gpt-6-luna-low-provider-adapter",
          "verified_publish": true,
          "levels_completed": 0,
          "actions": 110,
          "resets": 2,
          "state": "GAME_OVER",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            0,
            0,
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            110,
            0,
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            22,
            42,
            26,
            115,
            36,
            31,
            8,
            40,
            41
          ]
        },
        {
          "session_id": "7e50ff33-c145-4e7f-9737-ef348b4b4d0c",
          "game_id": "su15",
          "game_version": "1944f8ab",
          "published_at": "2026-09-16T19:22:28.111149Z",
          "score": 0,
          "model": "gpt-6-luna",
          "runner": "benchmark_agent",
          "config": "openai-gpt-6-luna-none",
          "verified_publish": true,
          "levels_completed": 0,
          "actions": 110,
          "resets": 4,
          "state": "GAME_OVER",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            0,
            0,
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            110,
            0,
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            22,
            42,
            26,
            115,
            36,
            31,
            8,
            40,
            41
          ]
        },
        {
          "session_id": "98b6e669-83f6-4f5f-8b78-94ec3c54c883",
          "game_id": "su15",
          "game_version": "1944f8ab",
          "published_at": "2026-09-16T16:06:56.708909Z",
          "score": 0,
          "model": "gpt-6-luna",
          "runner": "benchmark_agent",
          "config": "openai-gpt-6-luna-xhigh",
          "verified_publish": true,
          "levels_completed": 0,
          "actions": 110,
          "resets": 2,
          "state": "GAME_OVER",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            0,
            0,
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            110,
            0,
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            22,
            42,
            26,
            115,
            36,
            31,
            8,
            40,
            41
          ]
        },
        {
          "session_id": "1d429c16-bf8a-4c47-8f70-628f84815597",
          "game_id": "su15",
          "game_version": "1944f8ab",
          "published_at": "2026-09-16T15:01:56.360851Z",
          "score": 0,
          "model": "gpt-6-luna",
          "runner": "benchmark_agent",
          "config": "openai-gpt-6-luna-high-provider-adapter",
          "verified_publish": true,
          "levels_completed": 0,
          "actions": 110,
          "resets": 2,
          "state": "GAME_OVER",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            0,
            0,
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            110,
            0,
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            22,
            42,
            26,
            115,
            36,
            31,
            8,
            40,
            41
          ]
        },
        {
          "session_id": "4e7063c2-c0c0-4f89-a9ae-f26d5de6f6d0",
          "game_id": "su15",
          "game_version": "1944f8ab",
          "published_at": "2026-09-16T14:46:23.749601Z",
          "score": 1.2789007794953096,
          "model": "gpt-6-1-sol",
          "runner": "benchmark_agent",
          "config": "openai-gpt-6-1-sol-low",
          "verified_publish": true,
          "levels_completed": 1,
          "actions": 125,
          "resets": 1,
          "state": "GAME_OVER",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            57.550535077288934,
            0,
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            29,
            96,
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            22,
            42,
            26,
            115,
            36,
            31,
            8,
            40,
            41
          ]
        },
        {
          "session_id": "5782167d-d5dc-4951-9cee-24c019c84405",
          "game_id": "su15",
          "game_version": "1944f8ab",
          "published_at": "2026-09-16T09:09:23.745753Z",
          "score": 0,
          "model": "gpt-6-luna",
          "runner": "benchmark_agent",
          "config": "openai-gpt-6-luna-low",
          "verified_publish": true,
          "levels_completed": 0,
          "actions": 110,
          "resets": 2,
          "state": "GAME_OVER",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            0,
            0,
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            110,
            0,
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            22,
            42,
            26,
            115,
            36,
            31,
            8,
            40,
            41
          ]
        },
        {
          "session_id": "545dc43c-5883-4f06-ad1c-d70bf9fb9423",
          "game_id": "su15",
          "game_version": "1944f8ab",
          "published_at": "2026-09-16T08:49:10.200653Z",
          "score": 0,
          "model": "gpt-6-luna",
          "runner": "benchmark_agent",
          "config": "openai-gpt-6-luna-medium",
          "verified_publish": true,
          "levels_completed": 0,
          "actions": 110,
          "resets": 2,
          "state": "GAME_OVER",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            0,
            0,
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            110,
            0,
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            22,
            42,
            26,
            115,
            36,
            31,
            8,
            40,
            41
          ]
        },
        {
          "session_id": "072038f7-fa9b-42b7-ab3a-3b0272314da9",
          "game_id": "su15",
          "game_version": "1944f8ab",
          "published_at": "2026-09-16T08:22:13.580852Z",
          "score": 0,
          "model": "gpt-6-luna",
          "runner": "benchmark_agent",
          "config": "openai-gpt-6-luna-high",
          "verified_publish": true,
          "levels_completed": 0,
          "actions": 110,
          "resets": 2,
          "state": "GAME_OVER",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            0,
            0,
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            110,
            0,
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            22,
            42,
            26,
            115,
            36,
            31,
            8,
            40,
            41
          ]
        },
        {
          "session_id": "25a175d7-b2b4-4a9d-999a-c79828db3ed0",
          "game_id": "su15",
          "game_version": "1944f8ab",
          "published_at": "2026-09-16T07:55:30.693521Z",
          "score": 0,
          "model": "gpt-6-luna",
          "runner": "benchmark_agent",
          "config": "openai-gpt-6-luna-xhigh-provider-adapter",
          "verified_publish": true,
          "levels_completed": 0,
          "actions": 110,
          "resets": 3,
          "state": "GAME_OVER",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            0,
            0,
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            110,
            0,
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            22,
            42,
            26,
            115,
            36,
            31,
            8,
            40,
            41
          ]
        },
        {
          "session_id": "85b98a75-8a82-4d21-95c1-f5580208a4ed",
          "game_id": "su15",
          "game_version": "1944f8ab",
          "published_at": "2026-09-16T07:07:22.572606Z",
          "score": 0,
          "model": "gpt-6-luna",
          "runner": "benchmark_agent",
          "config": "openai-gpt-6-luna-max-provider-adapter",
          "verified_publish": true,
          "levels_completed": 0,
          "actions": 110,
          "resets": 3,
          "state": "GAME_OVER",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            0,
            0,
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            110,
            0,
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            22,
            42,
            26,
            115,
            36,
            31,
            8,
            40,
            41
          ]
        },
        {
          "session_id": "3e9d63dc-a057-48c0-8911-232759fff852",
          "game_id": "su15",
          "game_version": "1944f8ab",
          "published_at": "2026-09-16T06:59:16.573096Z",
          "score": 0,
          "model": "gpt-6-luna",
          "runner": "benchmark_agent",
          "config": "openai-gpt-6-luna-max",
          "verified_publish": true,
          "levels_completed": 0,
          "actions": 110,
          "resets": 2,
          "state": "GAME_OVER",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            0,
            0,
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            110,
            0,
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            22,
            42,
            26,
            115,
            36,
            31,
            8,
            40,
            41
          ]
        },
        {
          "session_id": "8087b5b4-010a-4918-94f4-8ea261aa230d",
          "game_id": "su15",
          "game_version": "1944f8ab",
          "published_at": "2026-09-16T06:11:18.47124Z",
          "score": 7.046038483327018,
          "model": "gpt-6-1-sol",
          "runner": "benchmark_agent",
          "config": "openai-gpt-6-1-sol-high",
          "verified_publish": true,
          "levels_completed": 4,
          "actions": 829,
          "resets": 20,
          "state": "GAME_OVER",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            57.550535077288934,
            37.051039697542535,
            55.18367346938776,
            4.967024217294634,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            29,
            69,
            35,
            516,
            180,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            22,
            42,
            26,
            115,
            36,
            31,
            8,
            40,
            41
          ]
        },
        {
          "session_id": "5d283c14-0ebe-4ee0-816c-298528dbb3d2",
          "game_id": "su15",
          "game_version": "1944f8ab",
          "published_at": "2026-09-16T05:51:47.332062Z",
          "score": 19.599393558168416,
          "model": "gpt-6-1-sol",
          "runner": "benchmark_agent",
          "config": "openai-gpt-6-1-sol-xhigh",
          "verified_publish": true,
          "levels_completed": 4,
          "actions": 391,
          "resets": 10,
          "state": "GAME_OVER",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            115,
            115,
            115,
            47.99317752939468,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            9,
            14,
            22,
            166,
            180,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            22,
            42,
            26,
            115,
            36,
            31,
            8,
            40,
            41
          ]
        },
        {
          "session_id": "183caf5c-fd38-4481-81b7-fcc5a9f04361",
          "game_id": "su15",
          "game_version": "1944f8ab",
          "published_at": "2026-09-16T05:14:32.082227Z",
          "score": 1.1533767551804317,
          "model": "gpt-6-1-sol",
          "runner": "benchmark_agent",
          "config": "openai-gpt-6-1-sol-medium",
          "verified_publish": true,
          "levels_completed": 2,
          "actions": 289,
          "resets": 4,
          "state": "GAME_OVER",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            11.11111111111111,
            20.39542143600416,
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            66,
            93,
            130,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            22,
            42,
            26,
            115,
            36,
            31,
            8,
            40,
            41
          ]
        },
        {
          "session_id": "77a207d2-e462-4bc3-b7e1-0aa172864e3d",
          "game_id": "su15",
          "game_version": "1944f8ab",
          "published_at": "2026-09-16T04:44:30.923138Z",
          "score": 100,
          "model": "gpt-6-1-sol",
          "runner": "benchmark_agent",
          "config": "openai-gpt-6-1-sol-high-provider-adapter",
          "verified_publish": true,
          "levels_completed": 9,
          "actions": 114,
          "resets": 0,
          "state": "WIN",
          "completed": true,
          "number_of_levels": 0,
          "level_scores": [
            115,
            115,
            115,
            115,
            115,
            115,
            115,
            115,
            115
          ],
          "level_actions": [
            9,
            10,
            17,
            35,
            8,
            9,
            6,
            10,
            10
          ],
          "level_baseline_actions": [
            22,
            42,
            26,
            115,
            36,
            31,
            8,
            40,
            41
          ]
        },
        {
          "session_id": "d5f37bff-9642-43e2-82fd-8976df0d9bbb",
          "game_id": "su15",
          "game_version": "1944f8ab",
          "published_at": "2026-09-16T04:13:55.597889Z",
          "score": 100,
          "model": "gpt-6-1-sol",
          "runner": "benchmark_agent",
          "config": "openai-gpt-6-1-sol-xhigh-provider-adapter",
          "verified_publish": true,
          "levels_completed": 9,
          "actions": 140,
          "resets": 2,
          "state": "WIN",
          "completed": true,
          "number_of_levels": 0,
          "level_scores": [
            115,
            115,
            115,
            115,
            115,
            115,
            115,
            115,
            115
          ],
          "level_actions": [
            8,
            15,
            16,
            36,
            9,
            18,
            6,
            12,
            20
          ],
          "level_baseline_actions": [
            22,
            42,
            26,
            115,
            36,
            31,
            8,
            40,
            41
          ]
        },
        {
          "session_id": "d7c0910b-fd0a-4513-86e4-2de5af94d11f",
          "game_id": "su15",
          "game_version": "1944f8ab",
          "published_at": "2026-09-16T02:43:12.788305Z",
          "score": 100,
          "model": "gpt-6-1-sol",
          "runner": "benchmark_agent",
          "config": "openai-gpt-6-1-sol-max-provider-adapter",
          "verified_publish": true,
          "levels_completed": 9,
          "actions": 145,
          "resets": 1,
          "state": "WIN",
          "completed": true,
          "number_of_levels": 0,
          "level_scores": [
            115,
            115,
            115,
            115,
            115,
            25.82639075517334,
            115,
            115,
            115
          ],
          "level_actions": [
            9,
            14,
            16,
            10,
            7,
            61,
            6,
            8,
            14
          ],
          "level_baseline_actions": [
            22,
            42,
            26,
            115,
            36,
            31,
            8,
            40,
            41
          ]
        },
        {
          "session_id": "847e1cfc-4d4f-47a2-b204-9eb54b030d8e",
          "game_id": "su15",
          "game_version": "1944f8ab",
          "published_at": "2026-09-16T01:17:38.496605Z",
          "score": 22.22222222222222,
          "model": "gpt-6-1-sol",
          "runner": "benchmark_agent",
          "config": "openai-gpt-6-1-sol-max",
          "verified_publish": true,
          "levels_completed": 4,
          "actions": 284,
          "resets": 4,
          "state": "GAME_OVER",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            115,
            115,
            115,
            115,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            10,
            11,
            21,
            62,
            180,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            22,
            42,
            26,
            115,
            36,
            31,
            8,
            40,
            41
          ]
        },
        {
          "session_id": "2449e701-fa4e-42e5-8292-b2a9e454d907",
          "game_id": "su15",
          "game_version": "1944f8ab",
          "published_at": "2026-09-12T01:34:41.994213Z",
          "score": 6.875959390684794,
          "model": "gemini-3.8-flash",
          "runner": "benchmark_agent",
          "config": "google-gemini-3-8-flash-medium",
          "verified_publish": true,
          "levels_completed": 3,
          "actions": 765,
          "resets": 19,
          "state": "GAME_OVER",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            115,
            91.11570247933885,
            4.062255874046031,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            17,
            44,
            129,
            575,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            22,
            42,
            26,
            115,
            36,
            31,
            8,
            40,
            41
          ]
        },
        {
          "session_id": "bbe4188a-f56a-4140-bcd5-f80d8a212135",
          "game_id": "su15",
          "game_version": "1944f8ab",
          "published_at": "2026-09-11T03:47:15.108365Z",
          "score": 12.077843523997371,
          "model": "gemini-3.8-flash",
          "runner": "benchmark_agent",
          "config": "google-gemini-3-8-flash-high",
          "verified_publish": true,
          "levels_completed": 3,
          "actions": 477,
          "resets": 28,
          "state": "GAME_OVER",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            115,
            41.75147928994083,
            115,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            13,
            65,
            22,
            377,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            22,
            42,
            26,
            115,
            36,
            31,
            8,
            40,
            41
          ]
        },
        {
          "session_id": "c41d37a4-a7a2-4e5a-a57b-641bbd2bdae3",
          "game_id": "su15",
          "game_version": "1944f8ab",
          "published_at": "2026-09-09T17:17:18.155537Z",
          "score": 1.1950617283950615,
          "model": "gemini-3.8-flash",
          "runner": "benchmark_agent",
          "config": "google-gemini-3-8-flash-low",
          "verified_publish": true,
          "levels_completed": 1,
          "actions": 41,
          "resets": 0,
          "state": "NOT_FINISHED",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            53.77777777777777,
            0,
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            30,
            11,
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            22,
            42,
            26,
            115,
            36,
            31,
            8,
            40,
            41
          ]
        },
        {
          "session_id": "3cd4e4fc-97fe-442b-ae2f-dea4593bc520",
          "game_id": "su15",
          "game_version": "1944f8ab",
          "published_at": "2026-08-11T06:20:43.07131Z",
          "score": 2.2222222222222223,
          "model": "grok-4.6",
          "runner": "benchmark_agent",
          "config": "xai-grok-4-6-xhigh",
          "verified_publish": true,
          "levels_completed": 1,
          "actions": 227,
          "resets": 2,
          "state": "GAME_OVER",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            115,
            0,
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            17,
            210,
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            22,
            42,
            26,
            115,
            36,
            31,
            8,
            40,
            41
          ]
        },
        {
          "session_id": "1a7d29a7-15c9-4d76-90d6-9629698c32c0",
          "game_id": "su15",
          "game_version": "1944f8ab",
          "published_at": "2026-07-21T05:30:52.266631Z",
          "score": 10.39015731292517,
          "model": "claude-opus-5",
          "runner": "benchmark_agent",
          "config": "anthropic-claude-opus-5-high",
          "verified_publish": true,
          "levels_completed": 3,
          "actions": 657,
          "resets": 162,
          "state": "GAME_OVER",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            39.51020408163265,
            115,
            66.015625,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            35,
            15,
            32,
            575,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            22,
            42,
            26,
            115,
            36,
            31,
            8,
            40,
            41
          ]
        },
        {
          "session_id": "594b27c6-3d4b-4e66-8746-9da3572b7bf2",
          "game_id": "su15",
          "game_version": "1944f8ab",
          "published_at": "2026-06-24T10:45:50.275248Z",
          "score": 0,
          "model": "gpt-5-6-luna",
          "runner": "benchmark_agent",
          "config": "openai-gpt-5-6-luna-low",
          "verified_publish": true,
          "levels_completed": 0,
          "actions": 110,
          "resets": 2,
          "state": "GAME_OVER",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            0,
            0,
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            110,
            0,
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            22,
            42,
            26,
            115,
            36,
            31,
            8,
            40,
            41
          ]
        },
        {
          "session_id": "0b02f150-889d-4de5-92b5-cbb0544a83a4",
          "game_id": "su15",
          "game_version": "1944f8ab",
          "published_at": "2026-06-24T10:14:27.988711Z",
          "score": 0,
          "model": "gpt-5-6-luna",
          "runner": "benchmark_agent",
          "config": "openai-gpt-5-6-luna-medium",
          "verified_publish": true,
          "levels_completed": 0,
          "actions": 110,
          "resets": 11,
          "state": "NOT_FINISHED",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            0,
            0,
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            110,
            0,
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            22,
            42,
            26,
            115,
            36,
            31,
            8,
            40,
            41
          ]
        },
        {
          "session_id": "29db9f57-f554-4ddd-bf72-b51a499356c0",
          "game_id": "su15",
          "game_version": "1944f8ab",
          "published_at": "2026-06-24T03:04:03.210787Z",
          "score": 0,
          "model": "gpt-5-6-luna",
          "runner": "benchmark_agent",
          "config": "openai-gpt-5-6-luna-high",
          "verified_publish": true,
          "levels_completed": 0,
          "actions": 110,
          "resets": 5,
          "state": "GAME_OVER",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            0,
            0,
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            110,
            0,
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            22,
            42,
            26,
            115,
            36,
            31,
            8,
            40,
            41
          ]
        },
        {
          "session_id": "6a6cfea3-37c0-4465-82d9-bf08605a3190",
          "game_id": "su15",
          "game_version": "1944f8ab",
          "published_at": "2026-06-23T23:42:49.569105Z",
          "score": 0,
          "model": "gpt-5-6-luna",
          "runner": "benchmark_agent",
          "config": "openai-gpt-5-6-luna-xhigh",
          "verified_publish": true,
          "levels_completed": 0,
          "actions": 110,
          "resets": 3,
          "state": "GAME_OVER",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            0,
            0,
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            110,
            0,
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            22,
            42,
            26,
            115,
            36,
            31,
            8,
            40,
            41
          ]
        },
        {
          "session_id": "ed1ac443-69d3-49de-89dc-a7d0fa6611ca",
          "game_id": "su15",
          "game_version": "1944f8ab",
          "published_at": "2026-06-23T01:39:20.935531Z",
          "score": 0,
          "model": "gpt-5-6-luna",
          "runner": "benchmark_agent",
          "config": "openai-gpt-5-6-luna-max",
          "verified_publish": true,
          "levels_completed": 0,
          "actions": 110,
          "resets": 3,
          "state": "NOT_FINISHED",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            0,
            0,
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            110,
            0,
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            22,
            42,
            26,
            115,
            36,
            31,
            8,
            40,
            41
          ]
        },
        {
          "session_id": "5016cd7d-46c9-43d7-ad86-d048f0327716",
          "game_id": "su15",
          "game_version": "1944f8ab",
          "published_at": "2026-06-20T11:50:30.503302Z",
          "score": 0,
          "model": "gpt-5-6-terra",
          "runner": "benchmark_agent",
          "config": "openai-gpt-5-6-terra-low",
          "verified_publish": true,
          "levels_completed": 0,
          "actions": 110,
          "resets": 1,
          "state": "GAME_OVER",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            0,
            0,
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            110,
            0,
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            22,
            42,
            26,
            115,
            36,
            31,
            8,
            40,
            41
          ]
        },
        {
          "session_id": "b2460081-be36-46b9-9cd6-78e3f63d054c",
          "game_id": "su15",
          "game_version": "1944f8ab",
          "published_at": "2026-06-20T09:02:13.997388Z",
          "score": 0,
          "model": "gpt-5-6-terra",
          "runner": "benchmark_agent",
          "config": "openai-gpt-5-6-terra-medium",
          "verified_publish": true,
          "levels_completed": 0,
          "actions": 110,
          "resets": 2,
          "state": "GAME_OVER",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            0,
            0,
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            110,
            0,
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            22,
            42,
            26,
            115,
            36,
            31,
            8,
            40,
            41
          ]
        },
        {
          "session_id": "5076b477-3a16-4c40-81c6-1b04c8034067",
          "game_id": "su15",
          "game_version": "1944f8ab",
          "published_at": "2026-06-20T08:12:11.915819Z",
          "score": 0,
          "model": "gpt-5-6-terra",
          "runner": "benchmark_agent",
          "config": "openai-gpt-5-6-terra-high",
          "verified_publish": true,
          "levels_completed": 0,
          "actions": 110,
          "resets": 3,
          "state": "GAME_OVER",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            0,
            0,
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            110,
            0,
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            22,
            42,
            26,
            115,
            36,
            31,
            8,
            40,
            41
          ]
        },
        {
          "session_id": "c00fbf32-52b8-4110-b410-e8a26a47030e",
          "game_id": "su15",
          "game_version": "1944f8ab",
          "published_at": "2026-06-20T00:40:36.82871Z",
          "score": 0,
          "model": "gpt-5-6-terra",
          "runner": "benchmark_agent",
          "config": "openai-gpt-5-6-terra-xhigh",
          "verified_publish": true,
          "levels_completed": 0,
          "actions": 110,
          "resets": 2,
          "state": "GAME_OVER",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            0,
            0,
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            110,
            0,
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            22,
            42,
            26,
            115,
            36,
            31,
            8,
            40,
            41
          ]
        },
        {
          "session_id": "fb89c72c-a1a6-4642-8be2-c51ebb375ef1",
          "game_id": "su15",
          "game_version": "1944f8ab",
          "published_at": "2026-06-18T21:43:51.268564Z",
          "score": 0,
          "model": "gpt-5-6-terra",
          "runner": "benchmark_agent",
          "config": "openai-gpt-5-6-terra-max",
          "verified_publish": true,
          "levels_completed": 0,
          "actions": 110,
          "resets": 4,
          "state": "GAME_OVER",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            0,
            0,
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            110,
            0,
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            22,
            42,
            26,
            115,
            36,
            31,
            8,
            40,
            41
          ]
        },
        {
          "session_id": "3a8d8e96-d380-4f8c-88a1-6a46fb29c2bc",
          "game_id": "su15",
          "game_version": "1944f8ab",
          "published_at": "2026-06-13T02:02:35.458195Z",
          "score": 0,
          "model": "gpt-5-6-sol",
          "runner": "benchmark_agent",
          "config": "openai-gpt-5-6-sol-max",
          "verified_publish": true,
          "levels_completed": 0,
          "actions": 110,
          "resets": 3,
          "state": "GAME_OVER",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            0,
            0,
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            110,
            0,
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            22,
            42,
            26,
            115,
            36,
            31,
            8,
            40,
            41
          ]
        },
        {
          "session_id": "fc0a811f-abf4-4970-9e1d-87da29edeab1",
          "game_id": "su15",
          "game_version": "1944f8ab",
          "published_at": "2026-06-11T13:08:21.15452Z",
          "score": 0,
          "model": "gpt-5-6-sol",
          "runner": "benchmark_agent",
          "config": "openai-gpt-5-6-sol-xhigh",
          "verified_publish": true,
          "levels_completed": 0,
          "actions": 110,
          "resets": 3,
          "state": "GAME_OVER",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            0,
            0,
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            110,
            0,
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            22,
            42,
            26,
            115,
            36,
            31,
            8,
            40,
            41
          ]
        },
        {
          "session_id": "6bcf6c78-f141-4534-9356-0f17046bed09",
          "game_id": "su15",
          "game_version": "1944f8ab",
          "published_at": "2026-06-10T10:02:43.707348Z",
          "score": 0,
          "model": "gpt-5-6-sol",
          "runner": "benchmark_agent",
          "config": "openai-gpt-5-6-sol-high",
          "verified_publish": true,
          "levels_completed": 0,
          "actions": 110,
          "resets": 2,
          "state": "GAME_OVER",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            0,
            0,
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            110,
            0,
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            22,
            42,
            26,
            115,
            36,
            31,
            8,
            40,
            41
          ]
        },
        {
          "session_id": "2b939228-ef61-43cf-823e-3808e4062ff7",
          "game_id": "su15",
          "game_version": "1944f8ab",
          "published_at": "2026-06-10T07:41:47.407582Z",
          "score": 0,
          "model": "gpt-5-6-sol",
          "runner": "benchmark_agent",
          "config": "openai-gpt-5-6-sol-medium",
          "verified_publish": true,
          "levels_completed": 0,
          "actions": 110,
          "resets": 2,
          "state": "GAME_OVER",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            0,
            0,
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            110,
            0,
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            22,
            42,
            26,
            115,
            36,
            31,
            8,
            40,
            41
          ]
        },
        {
          "session_id": "22f4736d-2c46-4ae4-b20b-8492883f7f7d",
          "game_id": "su15",
          "game_version": "1944f8ab",
          "published_at": "2026-06-09T21:25:45.851315Z",
          "score": 0,
          "model": "gpt-5-6-sol",
          "runner": "benchmark_agent",
          "config": "openai-gpt-5-6-sol-low",
          "verified_publish": true,
          "levels_completed": 0,
          "actions": 110,
          "resets": 2,
          "state": "GAME_OVER",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            0,
            0,
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            110,
            0,
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            22,
            42,
            26,
            115,
            36,
            31,
            8,
            40,
            41
          ]
        },
        {
          "session_id": "e4cfd8a1-4efc-44bf-8b25-82a086977e1f",
          "game_id": "su15",
          "game_version": "1944f8ab",
          "published_at": "2026-05-29T19:39:10.003583Z",
          "score": 1.1950617283950615,
          "model": "claude-opus-4-8",
          "runner": "benchmark_agent",
          "config": "anthropic-opus-4-8-high",
          "verified_publish": true,
          "levels_completed": 1,
          "actions": 240,
          "resets": 11,
          "state": "GAME_OVER",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            53.77777777777777,
            0,
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            30,
            210,
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            22,
            42,
            26,
            115,
            36,
            31,
            8,
            40,
            41
          ]
        },
        {
          "session_id": "a81ab0e9-f46a-452b-8f38-4c8d485e74f0",
          "game_id": "su15",
          "game_version": "1944f8ab",
          "published_at": "2026-05-01T05:07:22.085832Z",
          "score": 0,
          "model": "claude-opus-4-7",
          "runner": "benchmark_agent",
          "config": "anthropic-opus-4-7-high",
          "verified_publish": true,
          "levels_completed": 0,
          "actions": 110,
          "resets": 3,
          "state": "NOT_FINISHED",
          "completed": false,
          "number_of_levels": 9,
          "level_scores": [
            0,
            0,
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            110,
            0,
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            22,
            42,
            26,
            115,
            36,
            31,
            8,
            40,
            41
          ]
        },
        {
          "session_id": "c3c08777-eda6-4c5b-b1fd-74bfd13f0c31",
          "game_id": "su15",
          "game_version": "1944f8ab",
          "published_at": "2026-04-30T23:22:03.8915Z",
          "score": 0,
          "model": "gpt-5.5-2026-04-23",
          "runner": "benchmark_agent",
          "config": "openai-gpt-5-5-2026-04-23-high",
          "verified_publish": true,
          "levels_completed": 0,
          "actions": 110,
          "resets": 2,
          "state": "GAME_OVER",
          "completed": false,
          "number_of_levels": 9,
          "level_scores": [
            0,
            0,
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            110,
            0,
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            22,
            42,
            26,
            115,
            36,
            31,
            8,
            40,
            41
          ]
        },
        {
          "session_id": "ab0dc3e1-31db-48ba-811b-67c6cbd2a0f6",
          "game_id": "su15",
          "game_version": "1944f8ab",
          "published_at": "2026-04-14T17:43:44.846849Z",
          "score": 0,
          "model": "grok-4.20-beta-0309-reasoning",
          "runner": "benchmark_agent",
          "config": "xai-grok-4-20-beta-0309-reasoning",
          "verified_publish": true,
          "levels_completed": 0,
          "actions": 19,
          "resets": 0,
          "state": "NOT_FINISHED",
          "completed": false,
          "number_of_levels": 9,
          "level_scores": [
            0,
            0,
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            19,
            0,
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            22,
            42,
            26,
            115,
            36,
            31,
            8,
            40,
            41
          ]
        },
        {
          "session_id": "b0348f14-4749-478e-9973-4474cc114398",
          "game_id": "su15",
          "game_version": "1944f8ab",
          "published_at": "2026-04-14T17:43:44.519356Z",
          "score": 0,
          "model": "gpt-5.4-2026-03-05",
          "runner": "benchmark_agent",
          "config": "openai-gpt-5-4-2026-03-05-high",
          "verified_publish": true,
          "levels_completed": 0,
          "actions": 110,
          "resets": 4,
          "state": "GAME_OVER",
          "completed": false,
          "number_of_levels": 9,
          "level_scores": [
            0,
            0,
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            110,
            0,
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            22,
            42,
            26,
            115,
            36,
            31,
            8,
            40,
            41
          ]
        },
        {
          "session_id": "2e1f5876-3a43-4b42-806a-6853232b1395",
          "game_id": "su15",
          "game_version": "1944f8ab",
          "published_at": "2026-04-14T17:43:43.764965Z",
          "score": 0,
          "model": "gemini-3.1-pro-preview",
          "runner": "benchmark_agent",
          "config": "google-gemini-3-1-pro-preview",
          "verified_publish": true,
          "levels_completed": 0,
          "actions": 110,
          "resets": 5,
          "state": "NOT_FINISHED",
          "completed": false,
          "number_of_levels": 9,
          "level_scores": [
            0,
            0,
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            110,
            0,
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            22,
            42,
            26,
            115,
            36,
            31,
            8,
            40,
            41
          ]
        },
        {
          "session_id": "8dbfe114-030e-4aa3-8fe8-5703fa5ce300",
          "game_id": "su15",
          "game_version": "1944f8ab",
          "published_at": "2026-04-14T17:43:43.439694Z",
          "score": 0,
          "model": "claude-opus-4-6",
          "runner": "benchmark_agent",
          "config": "anthropic-opus-4-6-max-effort",
          "verified_publish": true,
          "levels_completed": 0,
          "actions": 110,
          "resets": 5,
          "state": "GAME_OVER",
          "completed": false,
          "number_of_levels": 9,
          "level_scores": [
            0,
            0,
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            110,
            0,
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            22,
            42,
            26,
            115,
            36,
            31,
            8,
            40,
            41
          ]
        }
      ]
    },
    {
      "game": "cd82",
      "source": "https://arcprize.org/tasks/cd82",
      "runs": [
        {
          "session_id": "fac349e3-0e46-4a0c-bfd6-5c20538b934f",
          "game_id": "cd82",
          "game_version": "fb555c5d",
          "published_at": "2026-09-25T07:48:53.01187Z",
          "score": 0,
          "model": "gpt-6-sol",
          "runner": "benchmark_agent",
          "config": "openai-gpt-6-sol-max",
          "verified_publish": true,
          "levels_completed": 0,
          "actions": 275,
          "resets": 2,
          "state": "GAME_OVER",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            275,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            55,
            8,
            41,
            21,
            23,
            23
          ]
        },
        {
          "session_id": "e75e1f6c-6020-4371-ba5c-521d89fadd91",
          "game_id": "cd82",
          "game_version": "fb555c5d",
          "published_at": "2026-09-25T06:50:01.018082Z",
          "score": 0,
          "model": "gpt-6-sol",
          "runner": "benchmark_agent",
          "config": "openai-gpt-6-sol-none-provider-adapter",
          "verified_publish": true,
          "levels_completed": 0,
          "actions": 275,
          "resets": 9,
          "state": "NOT_FINISHED",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            275,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            55,
            8,
            41,
            21,
            23,
            23
          ]
        },
        {
          "session_id": "06ebed49-e89f-419b-ad06-237d3e334365",
          "game_id": "cd82",
          "game_version": "fb555c5d",
          "published_at": "2026-09-25T06:31:53.978Z",
          "score": 0,
          "model": "gpt-6-sol",
          "runner": "benchmark_agent",
          "config": "openai-gpt-6-sol-none",
          "verified_publish": true,
          "levels_completed": 0,
          "actions": 275,
          "resets": 2,
          "state": "GAME_OVER",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            275,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            55,
            8,
            41,
            21,
            23,
            23
          ]
        },
        {
          "session_id": "a359b9ac-bd7e-4dd6-b950-29494de3c22f",
          "game_id": "cd82",
          "game_version": "fb555c5d",
          "published_at": "2026-09-25T06:22:33.46687Z",
          "score": 0.8523527754296985,
          "model": "gpt-6-sol",
          "runner": "benchmark_agent",
          "config": "openai-gpt-6-sol-low-provider-adapter",
          "verified_publish": true,
          "levels_completed": 1,
          "actions": 170,
          "resets": 1,
          "state": "NOT_FINISHED",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            17.899408284023668,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            130,
            40,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            55,
            8,
            41,
            21,
            23,
            23
          ]
        },
        {
          "session_id": "c0936a1e-0bbc-4037-a504-35ae1348d226",
          "game_id": "cd82",
          "game_version": "fb555c5d",
          "published_at": "2026-09-25T06:08:10.78264Z",
          "score": 0,
          "model": "gpt-6-sol",
          "runner": "benchmark_agent",
          "config": "openai-gpt-6-sol-low",
          "verified_publish": true,
          "levels_completed": 0,
          "actions": 275,
          "resets": 2,
          "state": "GAME_OVER",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            275,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            55,
            8,
            41,
            21,
            23,
            23
          ]
        },
        {
          "session_id": "1aced1e5-9f73-484c-94eb-ebae5330c9e7",
          "game_id": "cd82",
          "game_version": "fb555c5d",
          "published_at": "2026-09-25T05:45:46.820573Z",
          "score": 1.7783656672545565,
          "model": "gpt-6-sol",
          "runner": "benchmark_agent",
          "config": "openai-gpt-6-sol-medium-provider-adapter",
          "verified_publish": true,
          "levels_completed": 1,
          "actions": 130,
          "resets": 2,
          "state": "NOT_FINISHED",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            37.345679012345684,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            90,
            40,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            55,
            8,
            41,
            21,
            23,
            23
          ]
        },
        {
          "session_id": "94500655-51d5-4ed5-ade8-5d66432d630e",
          "game_id": "cd82",
          "game_version": "fb555c5d",
          "published_at": "2026-09-25T05:35:27.922597Z",
          "score": 0,
          "model": "gpt-6-sol",
          "runner": "benchmark_agent",
          "config": "openai-gpt-6-sol-medium",
          "verified_publish": true,
          "levels_completed": 0,
          "actions": 275,
          "resets": 2,
          "state": "GAME_OVER",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            275,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            55,
            8,
            41,
            21,
            23,
            23
          ]
        },
        {
          "session_id": "887eb0e9-a2e6-40cf-968f-f8e245081b51",
          "game_id": "cd82",
          "game_version": "fb555c5d",
          "published_at": "2026-09-25T05:27:04.4874Z",
          "score": 0,
          "model": "gpt-6-sol",
          "runner": "benchmark_agent",
          "config": "openai-gpt-6-sol-xhigh",
          "verified_publish": true,
          "levels_completed": 0,
          "actions": 275,
          "resets": 3,
          "state": "GAME_OVER",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            275,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            55,
            8,
            41,
            21,
            23,
            23
          ]
        },
        {
          "session_id": "6307a0b4-d9ce-47d0-bc72-12313a01d989",
          "game_id": "cd82",
          "game_version": "fb555c5d",
          "published_at": "2026-09-25T05:26:54.864506Z",
          "score": 6.2371760396708416,
          "model": "gpt-6-sol",
          "runner": "benchmark_agent",
          "config": "openai-gpt-6-sol-high-provider-adapter",
          "verified_publish": true,
          "levels_completed": 2,
          "actions": 292,
          "resets": 8,
          "state": "NOT_FINISHED",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            55.24105186267347,
            37.8698224852071,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            74,
            13,
            205,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            55,
            8,
            41,
            21,
            23,
            23
          ]
        },
        {
          "session_id": "08786213-8166-4873-91d7-10884cacaad6",
          "game_id": "cd82",
          "game_version": "fb555c5d",
          "published_at": "2026-09-25T05:04:58.042325Z",
          "score": 0,
          "model": "gpt-6-sol",
          "runner": "benchmark_agent",
          "config": "openai-gpt-6-sol-high",
          "verified_publish": true,
          "levels_completed": 0,
          "actions": 275,
          "resets": 3,
          "state": "GAME_OVER",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            275,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            55,
            8,
            41,
            21,
            23,
            23
          ]
        },
        {
          "session_id": "00d9ed8f-eca5-429c-92a0-60938284de3e",
          "game_id": "cd82",
          "game_version": "fb555c5d",
          "published_at": "2026-09-25T04:11:07.391154Z",
          "score": 3.1152166749052568,
          "model": "gpt-6-sol",
          "runner": "benchmark_agent",
          "config": "openai-gpt-6-sol-xhigh-provider-adapter",
          "verified_publish": true,
          "levels_completed": 1,
          "actions": 108,
          "resets": 1,
          "state": "NOT_FINISHED",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            65.41955017301039,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            68,
            40,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            55,
            8,
            41,
            21,
            23,
            23
          ]
        },
        {
          "session_id": "a0a826f8-03ac-4f22-9e8e-af553d55cdf4",
          "game_id": "cd82",
          "game_version": "fb555c5d",
          "published_at": "2026-09-21T07:58:25.849609Z",
          "score": 23.56718192627824,
          "model": "gemini-3.8-flash",
          "runner": "benchmark_agent",
          "config": "google-gemini-3-8-flash-low-provider-adapter",
          "verified_publish": true,
          "levels_completed": 3,
          "actions": 213,
          "resets": 15,
          "state": "NOT_FINISHED",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            115,
            115,
            49.97027348394767,
            0,
            0,
            0
          ],
          "level_actions": [
            44,
            6,
            58,
            105,
            0,
            0
          ],
          "level_baseline_actions": [
            55,
            8,
            41,
            21,
            23,
            23
          ]
        },
        {
          "session_id": "8e6efec2-10b2-4b50-922d-8371ff0e0f6c",
          "game_id": "cd82",
          "game_version": "fb555c5d",
          "published_at": "2026-09-20T02:41:54.046982Z",
          "score": 47.61904761904761,
          "model": "gemini-3.8-flash",
          "runner": "benchmark_agent",
          "config": "google-gemini-3-8-flash-medium-provider-adapter",
          "verified_publish": true,
          "levels_completed": 4,
          "actions": 169,
          "resets": 8,
          "state": "NOT_FINISHED",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            115,
            115,
            115,
            115,
            0,
            0
          ],
          "level_actions": [
            14,
            6,
            20,
            14,
            115,
            0
          ],
          "level_baseline_actions": [
            55,
            8,
            41,
            21,
            23,
            23
          ]
        },
        {
          "session_id": "84290070-f69a-4e1d-8d28-4a086eb808fa",
          "game_id": "cd82",
          "game_version": "fb555c5d",
          "published_at": "2026-09-17T16:09:57.444358Z",
          "score": 39.93055507041521,
          "model": "gemini-3.8-flash",
          "runner": "benchmark_agent",
          "config": "google-gemini-3-8-flash-high-provider-adapter",
          "verified_publish": true,
          "levels_completed": 4,
          "actions": 207,
          "resets": 5,
          "state": "NOT_FINISHED",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            115,
            115,
            110.51939513477976,
            40.49586776859504,
            0,
            0
          ],
          "level_actions": [
            14,
            6,
            39,
            33,
            115,
            0
          ],
          "level_baseline_actions": [
            55,
            8,
            41,
            21,
            23,
            23
          ]
        },
        {
          "session_id": "a979ac5f-42d1-4b1e-bb3c-a556a80bfd70",
          "game_id": "cd82",
          "game_version": "fb555c5d",
          "published_at": "2026-09-16T19:54:55.15051Z",
          "score": 0,
          "model": "gpt-6-luna",
          "runner": "benchmark_agent",
          "config": "openai-gpt-6-luna-medium-provider-adapter",
          "verified_publish": true,
          "levels_completed": 0,
          "actions": 275,
          "resets": 2,
          "state": "GAME_OVER",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            275,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            55,
            8,
            41,
            21,
            23,
            23
          ]
        },
        {
          "session_id": "781f4e7b-6f9f-4aa5-948d-4c4662dc87c6",
          "game_id": "cd82",
          "game_version": "fb555c5d",
          "published_at": "2026-09-16T19:35:59.980596Z",
          "score": 0,
          "model": "gpt-6-luna",
          "runner": "benchmark_agent",
          "config": "openai-gpt-6-luna-low-provider-adapter",
          "verified_publish": true,
          "levels_completed": 0,
          "actions": 275,
          "resets": 2,
          "state": "GAME_OVER",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            275,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            55,
            8,
            41,
            21,
            23,
            23
          ]
        },
        {
          "session_id": "29eedb2f-1515-4b0f-912f-93e1f01d5f8f",
          "game_id": "cd82",
          "game_version": "fb555c5d",
          "published_at": "2026-09-16T19:35:18.857922Z",
          "score": 0,
          "model": "gpt-6-luna",
          "runner": "benchmark_agent",
          "config": "openai-gpt-6-luna-low",
          "verified_publish": true,
          "levels_completed": 0,
          "actions": 275,
          "resets": 2,
          "state": "GAME_OVER",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            275,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            55,
            8,
            41,
            21,
            23,
            23
          ]
        },
        {
          "session_id": "7c47293b-beba-44a7-b858-a5522b5dbf0d",
          "game_id": "cd82",
          "game_version": "fb555c5d",
          "published_at": "2026-09-16T19:33:30.584993Z",
          "score": 0,
          "model": "gpt-6-luna",
          "runner": "benchmark_agent",
          "config": "openai-gpt-6-luna-none-provider-adapter",
          "verified_publish": true,
          "levels_completed": 0,
          "actions": 275,
          "resets": 2,
          "state": "GAME_OVER",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            275,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            55,
            8,
            41,
            21,
            23,
            23
          ]
        },
        {
          "session_id": "128ac7f6-9bfc-4a19-a0f9-8295900fa03e",
          "game_id": "cd82",
          "game_version": "fb555c5d",
          "published_at": "2026-09-16T17:36:55.775229Z",
          "score": 2.778696355085244,
          "model": "gpt-6-luna",
          "runner": "benchmark_agent",
          "config": "openai-gpt-6-luna-none",
          "verified_publish": true,
          "levels_completed": 1,
          "actions": 112,
          "resets": 0,
          "state": "NOT_FINISHED",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            58.35262345679012,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            72,
            40,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            55,
            8,
            41,
            21,
            23,
            23
          ]
        },
        {
          "session_id": "0457fc78-070e-4dbe-955d-b32b1f298ed2",
          "game_id": "cd82",
          "game_version": "fb555c5d",
          "published_at": "2026-09-16T15:23:57.238038Z",
          "score": 0,
          "model": "gpt-6-luna",
          "runner": "benchmark_agent",
          "config": "openai-gpt-6-luna-medium",
          "verified_publish": true,
          "levels_completed": 0,
          "actions": 275,
          "resets": 2,
          "state": "GAME_OVER",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            275,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            55,
            8,
            41,
            21,
            23,
            23
          ]
        },
        {
          "session_id": "502a424f-38df-4dfa-bf4e-9a6605d837c6",
          "game_id": "cd82",
          "game_version": "fb555c5d",
          "published_at": "2026-09-16T09:52:54.81372Z",
          "score": 0,
          "model": "gpt-6-luna",
          "runner": "benchmark_agent",
          "config": "openai-gpt-6-luna-max",
          "verified_publish": true,
          "levels_completed": 0,
          "actions": 275,
          "resets": 8,
          "state": "NOT_FINISHED",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            275,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            55,
            8,
            41,
            21,
            23,
            23
          ]
        },
        {
          "session_id": "56fe3d34-816b-46ac-ad4a-bf60ba9b6430",
          "game_id": "cd82",
          "game_version": "fb555c5d",
          "published_at": "2026-09-16T09:26:27.582749Z",
          "score": 0,
          "model": "gpt-6-luna",
          "runner": "benchmark_agent",
          "config": "openai-gpt-6-luna-high",
          "verified_publish": true,
          "levels_completed": 0,
          "actions": 275,
          "resets": 2,
          "state": "GAME_OVER",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            275,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            55,
            8,
            41,
            21,
            23,
            23
          ]
        },
        {
          "session_id": "5f99326c-ac79-4205-ae35-856d2a90c984",
          "game_id": "cd82",
          "game_version": "fb555c5d",
          "published_at": "2026-09-16T09:16:09.185465Z",
          "score": 0,
          "model": "gpt-6-luna",
          "runner": "benchmark_agent",
          "config": "openai-gpt-6-luna-high-provider-adapter",
          "verified_publish": true,
          "levels_completed": 0,
          "actions": 275,
          "resets": 2,
          "state": "GAME_OVER",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            275,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            55,
            8,
            41,
            21,
            23,
            23
          ]
        },
        {
          "session_id": "b8bbd097-e113-47b1-9362-3aca13250b3a",
          "game_id": "cd82",
          "game_version": "fb555c5d",
          "published_at": "2026-09-16T08:40:51.28541Z",
          "score": 0,
          "model": "gpt-6-luna",
          "runner": "benchmark_agent",
          "config": "openai-gpt-6-luna-xhigh",
          "verified_publish": true,
          "levels_completed": 0,
          "actions": 275,
          "resets": 3,
          "state": "GAME_OVER",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            275,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            55,
            8,
            41,
            21,
            23,
            23
          ]
        },
        {
          "session_id": "9d633904-05df-4df8-adfd-50162319bf50",
          "game_id": "cd82",
          "game_version": "fb555c5d",
          "published_at": "2026-09-16T08:36:28.730063Z",
          "score": 0,
          "model": "gpt-6-luna",
          "runner": "benchmark_agent",
          "config": "openai-gpt-6-luna-xhigh-provider-adapter",
          "verified_publish": true,
          "levels_completed": 0,
          "actions": 275,
          "resets": 2,
          "state": "GAME_OVER",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            275,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            55,
            8,
            41,
            21,
            23,
            23
          ]
        },
        {
          "session_id": "36e81c09-158a-4efd-9abc-d19278ea19c0",
          "game_id": "cd82",
          "game_version": "fb555c5d",
          "published_at": "2026-09-16T08:34:04.635352Z",
          "score": 0,
          "model": "gpt-6-luna",
          "runner": "benchmark_agent",
          "config": "openai-gpt-6-luna-max-provider-adapter",
          "verified_publish": true,
          "levels_completed": 0,
          "actions": 275,
          "resets": 3,
          "state": "GAME_OVER",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            275,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            55,
            8,
            41,
            21,
            23,
            23
          ]
        },
        {
          "session_id": "8735013a-4bc0-4fc0-a096-3a76c3b2c14a",
          "game_id": "cd82",
          "game_version": "fb555c5d",
          "published_at": "2026-09-16T06:59:41.560879Z",
          "score": 0,
          "model": "gpt-6-1-sol",
          "runner": "benchmark_agent",
          "config": "openai-gpt-6-1-sol-low",
          "verified_publish": true,
          "levels_completed": 0,
          "actions": 275,
          "resets": 3,
          "state": "NOT_FINISHED",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            275,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            55,
            8,
            41,
            21,
            23,
            23
          ]
        },
        {
          "session_id": "f992fe5a-5d86-474d-904b-32a8f1be40db",
          "game_id": "cd82",
          "game_version": "fb555c5d",
          "published_at": "2026-09-16T05:40:39.173065Z",
          "score": 0,
          "model": "gpt-6-1-sol",
          "runner": "benchmark_agent",
          "config": "openai-gpt-6-1-sol-medium",
          "verified_publish": true,
          "levels_completed": 0,
          "actions": 275,
          "resets": 5,
          "state": "NOT_FINISHED",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            275,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            55,
            8,
            41,
            21,
            23,
            23
          ]
        },
        {
          "session_id": "7822fba0-795f-4b07-afc8-69075b900fee",
          "game_id": "cd82",
          "game_version": "fb555c5d",
          "published_at": "2026-09-16T04:34:48.069462Z",
          "score": 50.89745685549882,
          "model": "gpt-6-1-sol",
          "runner": "benchmark_agent",
          "config": "openai-gpt-6-1-sol-xhigh",
          "verified_publish": true,
          "levels_completed": 5,
          "actions": 309,
          "resets": 5,
          "state": "NOT_FINISHED",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            71.59763313609467,
            9.467455621301776,
            115,
            14.578512396694215,
            115,
            0
          ],
          "level_actions": [
            65,
            26,
            27,
            55,
            21,
            115
          ],
          "level_baseline_actions": [
            55,
            8,
            41,
            21,
            23,
            23
          ]
        },
        {
          "session_id": "dbb8cc43-9776-4300-8121-ad8f4d2cf72b",
          "game_id": "cd82",
          "game_version": "fb555c5d",
          "published_at": "2026-09-16T04:27:58.822492Z",
          "score": 100,
          "model": "gpt-6-1-sol",
          "runner": "benchmark_agent",
          "config": "openai-gpt-6-1-sol-high-provider-adapter",
          "verified_publish": true,
          "levels_completed": 6,
          "actions": 86,
          "resets": 1,
          "state": "WIN",
          "completed": true,
          "number_of_levels": 0,
          "level_scores": [
            115,
            115,
            115,
            115,
            115,
            115
          ],
          "level_actions": [
            15,
            6,
            19,
            14,
            16,
            16
          ],
          "level_baseline_actions": [
            55,
            8,
            41,
            21,
            23,
            23
          ]
        },
        {
          "session_id": "fa0d5cfa-d4a3-4050-b63a-7f45cad32c57",
          "game_id": "cd82",
          "game_version": "fb555c5d",
          "published_at": "2026-09-16T04:23:43.610376Z",
          "score": 7.1897925035349965,
          "model": "gpt-6-1-sol",
          "runner": "benchmark_agent",
          "config": "openai-gpt-6-1-sol-high",
          "verified_publish": true,
          "levels_completed": 3,
          "actions": 295,
          "resets": 5,
          "state": "GAME_OVER",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            55.24105186267347,
            22.145328719723185,
            17.15131109070503,
            0,
            0,
            0
          ],
          "level_actions": [
            74,
            17,
            99,
            105,
            0,
            0
          ],
          "level_baseline_actions": [
            55,
            8,
            41,
            21,
            23,
            23
          ]
        },
        {
          "session_id": "dcbc0335-2312-4193-a67d-cb39f5eb9d4c",
          "game_id": "cd82",
          "game_version": "fb555c5d",
          "published_at": "2026-09-16T03:16:31.34122Z",
          "score": 100,
          "model": "gpt-6-1-sol",
          "runner": "benchmark_agent",
          "config": "openai-gpt-6-1-sol-xhigh-provider-adapter",
          "verified_publish": true,
          "levels_completed": 6,
          "actions": 102,
          "resets": 2,
          "state": "WIN",
          "completed": true,
          "number_of_levels": 0,
          "level_scores": [
            115,
            115,
            115,
            115,
            115,
            115
          ],
          "level_actions": [
            31,
            6,
            17,
            16,
            16,
            16
          ],
          "level_baseline_actions": [
            55,
            8,
            41,
            21,
            23,
            23
          ]
        },
        {
          "session_id": "11b6a9e0-f851-4437-8355-630b0395633a",
          "game_id": "cd82",
          "game_version": "fb555c5d",
          "published_at": "2026-09-16T03:14:10.78204Z",
          "score": 100,
          "model": "gpt-6-1-sol",
          "runner": "benchmark_agent",
          "config": "openai-gpt-6-1-sol-max-provider-adapter",
          "verified_publish": true,
          "levels_completed": 6,
          "actions": 93,
          "resets": 0,
          "state": "WIN",
          "completed": true,
          "number_of_levels": 0,
          "level_scores": [
            115,
            115,
            115,
            115,
            115,
            115
          ],
          "level_actions": [
            21,
            6,
            18,
            16,
            16,
            16
          ],
          "level_baseline_actions": [
            55,
            8,
            41,
            21,
            23,
            23
          ]
        },
        {
          "session_id": "efe7acee-c49c-4f9d-aa15-a93265514aaf",
          "game_id": "cd82",
          "game_version": "fb555c5d",
          "published_at": "2026-09-15T21:20:48.929107Z",
          "score": 62.63988095238095,
          "model": "gpt-6-1-sol",
          "runner": "benchmark_agent",
          "config": "openai-gpt-6-1-sol-max",
          "verified_publish": true,
          "levels_completed": 5,
          "actions": 163,
          "resets": 2,
          "state": "NOT_FINISHED",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            115,
            100,
            105.06249999999999,
            27.5625,
            115,
            0
          ],
          "level_actions": [
            26,
            8,
            40,
            40,
            21,
            28
          ],
          "level_baseline_actions": [
            55,
            8,
            41,
            21,
            23,
            23
          ]
        },
        {
          "session_id": "ddccab7c-96d3-47ba-ab6c-9bec54c5e721",
          "game_id": "cd82",
          "game_version": "fb555c5d",
          "published_at": "2026-09-11T05:45:32.814266Z",
          "score": 1.5136501159252507,
          "model": "gemini-3.8-flash",
          "runner": "benchmark_agent",
          "config": "google-gemini-3-8-flash-low",
          "verified_publish": true,
          "levels_completed": 2,
          "actions": 465,
          "resets": 46,
          "state": "NOT_FINISHED",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            5.3403714426947255,
            13.223140495867769,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            238,
            22,
            205,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            55,
            8,
            41,
            21,
            23,
            23
          ]
        },
        {
          "session_id": "dc5c3df8-9b72-494f-a887-068d2688c00b",
          "game_id": "cd82",
          "game_version": "fb555c5d",
          "published_at": "2026-09-09T16:59:10.468609Z",
          "score": 4.761904761904762,
          "model": "gemini-3.8-flash",
          "runner": "benchmark_agent",
          "config": "google-gemini-3-8-flash-medium",
          "verified_publish": true,
          "levels_completed": 1,
          "actions": 55,
          "resets": 4,
          "state": "NOT_FINISHED",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            115,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            39,
            16,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            55,
            8,
            41,
            21,
            23,
            23
          ]
        },
        {
          "session_id": "8f312b9a-9b94-4efd-8b59-02109c1dcb81",
          "game_id": "cd82",
          "game_version": "fb555c5d",
          "published_at": "2026-09-09T16:02:58.540029Z",
          "score": 4.761904761904762,
          "model": "gemini-3.8-flash",
          "runner": "benchmark_agent",
          "config": "google-gemini-3-8-flash-high",
          "verified_publish": true,
          "levels_completed": 1,
          "actions": 60,
          "resets": 1,
          "state": "NOT_FINISHED",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            115,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            20,
            40,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            55,
            8,
            41,
            21,
            23,
            23
          ]
        },
        {
          "session_id": "88346fa1-6bbb-4f66-89a7-5831ad4cc7f1",
          "game_id": "cd82",
          "game_version": "fb555c5d",
          "published_at": "2026-08-11T09:21:47.712639Z",
          "score": 0,
          "model": "grok-4.6",
          "runner": "benchmark_agent",
          "config": "xai-grok-4-6-xhigh",
          "verified_publish": true,
          "levels_completed": 0,
          "actions": 275,
          "resets": 3,
          "state": "GAME_OVER",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            275,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            55,
            8,
            41,
            21,
            23,
            23
          ]
        },
        {
          "session_id": "7944ede0-0f14-498f-9deb-fcc50a44e90f",
          "game_id": "cd82",
          "game_version": "fb555c5d",
          "published_at": "2026-07-21T03:56:21.271863Z",
          "score": 0.5919116496039574,
          "model": "claude-opus-5",
          "runner": "benchmark_agent",
          "config": "anthropic-claude-opus-5-high",
          "verified_publish": true,
          "levels_completed": 1,
          "actions": 196,
          "resets": 3,
          "state": "NOT_FINISHED",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            12.430144641683105,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            156,
            40,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            55,
            8,
            41,
            21,
            23,
            23
          ]
        },
        {
          "session_id": "9b4de4b7-b636-474c-b1a5-9940ff76b991",
          "game_id": "cd82",
          "game_version": "fb555c5d",
          "published_at": "2026-06-24T10:58:14.506026Z",
          "score": 0,
          "model": "gpt-5-6-luna",
          "runner": "benchmark_agent",
          "config": "openai-gpt-5-6-luna-low",
          "verified_publish": true,
          "levels_completed": 0,
          "actions": 275,
          "resets": 5,
          "state": "GAME_OVER",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            275,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            55,
            8,
            41,
            21,
            23,
            23
          ]
        },
        {
          "session_id": "6833280b-b770-4335-8c2a-dd2bf1b3a6c6",
          "game_id": "cd82",
          "game_version": "fb555c5d",
          "published_at": "2026-06-24T10:43:40.937158Z",
          "score": 0,
          "model": "gpt-5-6-luna",
          "runner": "benchmark_agent",
          "config": "openai-gpt-5-6-luna-medium",
          "verified_publish": true,
          "levels_completed": 0,
          "actions": 275,
          "resets": 2,
          "state": "GAME_OVER",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            275,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            55,
            8,
            41,
            21,
            23,
            23
          ]
        },
        {
          "session_id": "9fdac759-fbab-4a00-9d30-a05525b86a2b",
          "game_id": "cd82",
          "game_version": "fb555c5d",
          "published_at": "2026-06-24T04:31:25.48379Z",
          "score": 0,
          "model": "gpt-5-6-luna",
          "runner": "benchmark_agent",
          "config": "openai-gpt-5-6-luna-high",
          "verified_publish": true,
          "levels_completed": 0,
          "actions": 275,
          "resets": 20,
          "state": "NOT_FINISHED",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            275,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            55,
            8,
            41,
            21,
            23,
            23
          ]
        },
        {
          "session_id": "9ffb9f92-cfd0-4e2a-bbb4-9b065ca5c11b",
          "game_id": "cd82",
          "game_version": "fb555c5d",
          "published_at": "2026-06-24T02:53:09.869674Z",
          "score": 0,
          "model": "gpt-5-6-luna",
          "runner": "benchmark_agent",
          "config": "openai-gpt-5-6-luna-xhigh",
          "verified_publish": true,
          "levels_completed": 0,
          "actions": 275,
          "resets": 37,
          "state": "NOT_FINISHED",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            275,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            55,
            8,
            41,
            21,
            23,
            23
          ]
        },
        {
          "session_id": "2c194e8d-9a23-442c-9338-c193dd16d234",
          "game_id": "cd82",
          "game_version": "fb555c5d",
          "published_at": "2026-06-23T04:59:03.805788Z",
          "score": 0,
          "model": "gpt-5-6-luna",
          "runner": "benchmark_agent",
          "config": "openai-gpt-5-6-luna-max",
          "verified_publish": true,
          "levels_completed": 0,
          "actions": 275,
          "resets": 55,
          "state": "NOT_FINISHED",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            275,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            55,
            8,
            41,
            21,
            23,
            23
          ]
        },
        {
          "session_id": "904a8532-1126-47bd-9347-791fd1578a95",
          "game_id": "cd82",
          "game_version": "fb555c5d",
          "published_at": "2026-06-20T11:57:49.116152Z",
          "score": 0,
          "model": "gpt-5-6-terra",
          "runner": "benchmark_agent",
          "config": "openai-gpt-5-6-terra-low",
          "verified_publish": true,
          "levels_completed": 0,
          "actions": 86,
          "resets": 0,
          "state": "NOT_FINISHED",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            86,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            55,
            8,
            41,
            21,
            23,
            23
          ]
        },
        {
          "session_id": "24ef7f12-d1b7-4230-9bb8-c95183b5d6af",
          "game_id": "cd82",
          "game_version": "fb555c5d",
          "published_at": "2026-06-20T09:39:01.628373Z",
          "score": 0,
          "model": "gpt-5-6-terra",
          "runner": "benchmark_agent",
          "config": "openai-gpt-5-6-terra-medium",
          "verified_publish": true,
          "levels_completed": 0,
          "actions": 275,
          "resets": 3,
          "state": "GAME_OVER",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            275,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            55,
            8,
            41,
            21,
            23,
            23
          ]
        },
        {
          "session_id": "bb835284-7797-4d77-b16d-ed3b9d1edcec",
          "game_id": "cd82",
          "game_version": "fb555c5d",
          "published_at": "2026-06-20T09:21:25.877839Z",
          "score": 0,
          "model": "gpt-5-6-terra",
          "runner": "benchmark_agent",
          "config": "openai-gpt-5-6-terra-high",
          "verified_publish": true,
          "levels_completed": 0,
          "actions": 275,
          "resets": 3,
          "state": "GAME_OVER",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            275,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            55,
            8,
            41,
            21,
            23,
            23
          ]
        },
        {
          "session_id": "410ea2ff-aef8-4c4f-a8d0-025259b32ace",
          "game_id": "cd82",
          "game_version": "fb555c5d",
          "published_at": "2026-06-20T04:22:15.449711Z",
          "score": 0,
          "model": "gpt-5-6-terra",
          "runner": "benchmark_agent",
          "config": "openai-gpt-5-6-terra-xhigh",
          "verified_publish": true,
          "levels_completed": 0,
          "actions": 275,
          "resets": 14,
          "state": "NOT_FINISHED",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            275,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            55,
            8,
            41,
            21,
            23,
            23
          ]
        },
        {
          "session_id": "0c55c081-0864-4012-aa01-740c1eef246a",
          "game_id": "cd82",
          "game_version": "fb555c5d",
          "published_at": "2026-06-19T00:15:56.724798Z",
          "score": 0,
          "model": "gpt-5-6-terra",
          "runner": "benchmark_agent",
          "config": "openai-gpt-5-6-terra-max",
          "verified_publish": true,
          "levels_completed": 0,
          "actions": 275,
          "resets": 41,
          "state": "NOT_FINISHED",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            275,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            55,
            8,
            41,
            21,
            23,
            23
          ]
        },
        {
          "session_id": "a98c9bc0-7a74-44f4-abc8-6f4052c80b5e",
          "game_id": "cd82",
          "game_version": "fb555c5d",
          "published_at": "2026-06-12T23:42:01.988243Z",
          "score": 4.761904761904762,
          "model": "gpt-5-6-sol",
          "runner": "benchmark_agent",
          "config": "openai-gpt-5-6-sol-max",
          "verified_publish": true,
          "levels_completed": 1,
          "actions": 52,
          "resets": 7,
          "state": "NOT_FINISHED",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            115,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            12,
            40,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            55,
            8,
            41,
            21,
            23,
            23
          ]
        },
        {
          "session_id": "dd3b45dd-2848-4697-b110-594878ef936f",
          "game_id": "cd82",
          "game_version": "fb555c5d",
          "published_at": "2026-06-11T15:24:37.960154Z",
          "score": 3.6293176882746048,
          "model": "gpt-5-6-sol",
          "runner": "benchmark_agent",
          "config": "openai-gpt-5-6-sol-xhigh",
          "verified_publish": true,
          "levels_completed": 1,
          "actions": 103,
          "resets": 8,
          "state": "NOT_FINISHED",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            76.2156714537667,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            63,
            40,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            55,
            8,
            41,
            21,
            23,
            23
          ]
        },
        {
          "session_id": "f52d6f0e-f750-44f9-9958-987653772cd5",
          "game_id": "cd82",
          "game_version": "fb555c5d",
          "published_at": "2026-06-10T14:02:08.681641Z",
          "score": 0,
          "model": "gpt-5-6-sol",
          "runner": "benchmark_agent",
          "config": "openai-gpt-5-6-sol-high",
          "verified_publish": true,
          "levels_completed": 0,
          "actions": 275,
          "resets": 11,
          "state": "NOT_FINISHED",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            275,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            55,
            8,
            41,
            21,
            23,
            23
          ]
        },
        {
          "session_id": "f4d31916-f6d4-4325-ba44-9d8830873ea1",
          "game_id": "cd82",
          "game_version": "fb555c5d",
          "published_at": "2026-06-10T08:56:32.665628Z",
          "score": 0,
          "model": "gpt-5-6-sol",
          "runner": "benchmark_agent",
          "config": "openai-gpt-5-6-sol-medium",
          "verified_publish": true,
          "levels_completed": 0,
          "actions": 275,
          "resets": 4,
          "state": "GAME_OVER",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            275,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            55,
            8,
            41,
            21,
            23,
            23
          ]
        },
        {
          "session_id": "51653e66-accf-4c37-b672-3b41ca887e15",
          "game_id": "cd82",
          "game_version": "fb555c5d",
          "published_at": "2026-06-09T22:16:20.825965Z",
          "score": 0,
          "model": "gpt-5-6-sol",
          "runner": "benchmark_agent",
          "config": "openai-gpt-5-6-sol-low",
          "verified_publish": true,
          "levels_completed": 0,
          "actions": 275,
          "resets": 3,
          "state": "GAME_OVER",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            275,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            55,
            8,
            41,
            21,
            23,
            23
          ]
        },
        {
          "session_id": "d82d4ed2-87d1-403b-9289-fb1519976c6f",
          "game_id": "cd82",
          "game_version": "fb555c5d",
          "published_at": "2026-05-29T19:14:07.207985Z",
          "score": 0,
          "model": "claude-opus-4-8",
          "runner": "benchmark_agent",
          "config": "anthropic-opus-4-8-high",
          "verified_publish": true,
          "levels_completed": 0,
          "actions": 275,
          "resets": 7,
          "state": "NOT_FINISHED",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            275,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            55,
            8,
            41,
            21,
            23,
            23
          ]
        },
        {
          "session_id": "67ca6a82-fc59-4bcb-8646-e66edd66d2eb",
          "game_id": "cd82",
          "game_version": "fb555c5d",
          "published_at": "2026-05-01T05:08:28.222324Z",
          "score": 0,
          "model": "claude-opus-4-7",
          "runner": "benchmark_agent",
          "config": "anthropic-opus-4-7-high",
          "verified_publish": true,
          "levels_completed": 0,
          "actions": 275,
          "resets": 5,
          "state": "NOT_FINISHED",
          "completed": false,
          "number_of_levels": 6,
          "level_scores": [
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            275,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            55,
            8,
            41,
            21,
            23,
            23
          ]
        },
        {
          "session_id": "76dd45e6-402d-4586-8c3a-fdc1a2b191d1",
          "game_id": "cd82",
          "game_version": "fb555c5d",
          "published_at": "2026-04-30T23:24:01.886174Z",
          "score": 0,
          "model": "gpt-5.5-2026-04-23",
          "runner": "benchmark_agent",
          "config": "openai-gpt-5-5-2026-04-23-high",
          "verified_publish": true,
          "levels_completed": 0,
          "actions": 275,
          "resets": 4,
          "state": "GAME_OVER",
          "completed": false,
          "number_of_levels": 6,
          "level_scores": [
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            275,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            55,
            8,
            41,
            21,
            23,
            23
          ]
        },
        {
          "session_id": "0ab177fb-ada5-4090-9846-66b093206246",
          "game_id": "cd82",
          "game_version": "fb555c5d",
          "published_at": "2026-04-14T17:44:20.550303Z",
          "score": 0,
          "model": "grok-4.20-beta-0309-reasoning",
          "runner": "benchmark_agent",
          "config": "xai-grok-4-20-beta-0309-reasoning",
          "verified_publish": true,
          "levels_completed": 0,
          "actions": 275,
          "resets": 11,
          "state": "NOT_FINISHED",
          "completed": false,
          "number_of_levels": 6,
          "level_scores": [
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            275,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            55,
            8,
            41,
            21,
            23,
            23
          ]
        },
        {
          "session_id": "8c4b6113-54b2-414c-acda-4deb19178f10",
          "game_id": "cd82",
          "game_version": "fb555c5d",
          "published_at": "2026-04-14T17:44:20.028464Z",
          "score": 0,
          "model": "gpt-5.4-2026-03-05",
          "runner": "benchmark_agent",
          "config": "openai-gpt-5-4-2026-03-05-high",
          "verified_publish": true,
          "levels_completed": 0,
          "actions": 275,
          "resets": 3,
          "state": "GAME_OVER",
          "completed": false,
          "number_of_levels": 6,
          "level_scores": [
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            275,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            55,
            8,
            41,
            21,
            23,
            23
          ]
        },
        {
          "session_id": "ce0cbe75-18ce-4bdc-9879-e0290861442e",
          "game_id": "cd82",
          "game_version": "fb555c5d",
          "published_at": "2026-04-14T17:44:18.626813Z",
          "score": 0,
          "model": "gemini-3.1-pro-preview",
          "runner": "benchmark_agent",
          "config": "google-gemini-3-1-pro-preview",
          "verified_publish": true,
          "levels_completed": 0,
          "actions": 275,
          "resets": 21,
          "state": "NOT_FINISHED",
          "completed": false,
          "number_of_levels": 6,
          "level_scores": [
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            275,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            55,
            8,
            41,
            21,
            23,
            23
          ]
        },
        {
          "session_id": "58810dec-bb07-4e21-a6ca-4ac5d5d88f0c",
          "game_id": "cd82",
          "game_version": "fb555c5d",
          "published_at": "2026-04-14T17:44:17.904841Z",
          "score": 0,
          "model": "claude-opus-4-6",
          "runner": "benchmark_agent",
          "config": "anthropic-opus-4-6-max-effort",
          "verified_publish": true,
          "levels_completed": 0,
          "actions": 275,
          "resets": 24,
          "state": "NOT_FINISHED",
          "completed": false,
          "number_of_levels": 6,
          "level_scores": [
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            275,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            55,
            8,
            41,
            21,
            23,
            23
          ]
        }
      ]
    },
    {
      "game": "ft09",
      "source": "https://arcprize.org/tasks/ft09",
      "runs": [
        {
          "session_id": "5014efb5-298d-48f2-8708-df102f3b199f",
          "game_id": "ft09",
          "game_version": "0d8bbf25",
          "published_at": "2026-09-25T10:23:42.382294Z",
          "score": 38.379825302902226,
          "model": "gpt-6-sol",
          "runner": "benchmark_agent",
          "config": "openai-gpt-6-sol-xhigh",
          "verified_publish": true,
          "levels_completed": 4,
          "actions": 410,
          "resets": 2,
          "state": "GAME_OVER",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            115,
            115,
            115,
            28.994082840236686,
            0,
            0
          ],
          "level_actions": [
            4,
            9,
            20,
            52,
            325,
            0
          ],
          "level_baseline_actions": [
            43,
            12,
            23,
            28,
            65,
            37
          ]
        },
        {
          "session_id": "354ca84a-9803-4ccb-9b2d-d1876a70fb83",
          "game_id": "ft09",
          "game_version": "0d8bbf25",
          "published_at": "2026-09-25T08:10:03.439154Z",
          "score": 96.74107142857143,
          "model": "gpt-6-sol",
          "runner": "benchmark_agent",
          "config": "openai-gpt-6-sol-max",
          "verified_publish": true,
          "levels_completed": 6,
          "actions": 160,
          "resets": 0,
          "state": "WIN",
          "completed": true,
          "number_of_levels": 0,
          "level_scores": [
            115,
            115,
            115,
            19.140625,
            115,
            115
          ],
          "level_actions": [
            4,
            7,
            14,
            64,
            56,
            15
          ],
          "level_baseline_actions": [
            43,
            12,
            23,
            28,
            65,
            37
          ]
        },
        {
          "session_id": "9e53dc16-eeb9-4205-ad33-4d25e6e70146",
          "game_id": "ft09",
          "game_version": "0d8bbf25",
          "published_at": "2026-09-25T06:40:39.698321Z",
          "score": 14.285714285714285,
          "model": "gpt-6-sol",
          "runner": "benchmark_agent",
          "config": "openai-gpt-6-sol-none-provider-adapter",
          "verified_publish": true,
          "levels_completed": 2,
          "actions": 148,
          "resets": 0,
          "state": "NOT_FINISHED",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            115,
            115,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            22,
            11,
            115,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            43,
            12,
            23,
            28,
            65,
            37
          ]
        },
        {
          "session_id": "60dbda0c-94dd-4a98-9cef-0121268916e3",
          "game_id": "ft09",
          "game_version": "0d8bbf25",
          "published_at": "2026-09-25T06:32:56.239937Z",
          "score": 0,
          "model": "gpt-6-sol",
          "runner": "benchmark_agent",
          "config": "openai-gpt-6-sol-none",
          "verified_publish": true,
          "levels_completed": 0,
          "actions": 215,
          "resets": 2,
          "state": "GAME_OVER",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            215,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            43,
            12,
            23,
            28,
            65,
            37
          ]
        },
        {
          "session_id": "0788e4c2-bd3e-4228-9f2f-eb9e5ade3380",
          "game_id": "ft09",
          "game_version": "0d8bbf25",
          "published_at": "2026-09-25T06:17:21.382843Z",
          "score": 28.57142857142857,
          "model": "gpt-6-sol",
          "runner": "benchmark_agent",
          "config": "openai-gpt-6-sol-low-provider-adapter",
          "verified_publish": true,
          "levels_completed": 3,
          "actions": 168,
          "resets": 1,
          "state": "NOT_FINISHED",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            115,
            115,
            115,
            0,
            0,
            0
          ],
          "level_actions": [
            6,
            8,
            14,
            140,
            0,
            0
          ],
          "level_baseline_actions": [
            43,
            12,
            23,
            28,
            65,
            37
          ]
        },
        {
          "session_id": "0e097531-bda0-4040-a727-2f7e8eff6061",
          "game_id": "ft09",
          "game_version": "0d8bbf25",
          "published_at": "2026-09-25T06:15:20.786727Z",
          "score": 0,
          "model": "gpt-6-sol",
          "runner": "benchmark_agent",
          "config": "openai-gpt-6-sol-low",
          "verified_publish": true,
          "levels_completed": 0,
          "actions": 215,
          "resets": 2,
          "state": "NOT_FINISHED",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            215,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            43,
            12,
            23,
            28,
            65,
            37
          ]
        },
        {
          "session_id": "d86f45e1-3bb5-4abf-ab6e-436ac5c1e4e3",
          "game_id": "ft09",
          "game_version": "0d8bbf25",
          "published_at": "2026-09-25T05:58:11.957099Z",
          "score": 14.755102040816325,
          "model": "gpt-6-sol",
          "runner": "benchmark_agent",
          "config": "openai-gpt-6-sol-medium-provider-adapter",
          "verified_publish": true,
          "levels_completed": 3,
          "actions": 203,
          "resets": 1,
          "state": "GAME_OVER",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            115,
            32.6530612244898,
            43.183673469387756,
            0,
            0,
            0
          ],
          "level_actions": [
            7,
            21,
            35,
            140,
            0,
            0
          ],
          "level_baseline_actions": [
            43,
            12,
            23,
            28,
            65,
            37
          ]
        },
        {
          "session_id": "15569026-1794-4186-a29a-4d00f8b2e637",
          "game_id": "ft09",
          "game_version": "0d8bbf25",
          "published_at": "2026-09-25T05:45:26.637766Z",
          "score": 14.285714285714285,
          "model": "gpt-6-sol",
          "runner": "benchmark_agent",
          "config": "openai-gpt-6-sol-medium",
          "verified_publish": true,
          "levels_completed": 2,
          "actions": 134,
          "resets": 0,
          "state": "NOT_FINISHED",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            115,
            115,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            9,
            10,
            115,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            43,
            12,
            23,
            28,
            65,
            37
          ]
        },
        {
          "session_id": "7a4896a2-0918-4af1-9217-54e490e3cdaf",
          "game_id": "ft09",
          "game_version": "0d8bbf25",
          "published_at": "2026-09-25T05:18:55.370177Z",
          "score": 67.62051520080956,
          "model": "gpt-6-sol",
          "runner": "benchmark_agent",
          "config": "openai-gpt-6-sol-xhigh-provider-adapter",
          "verified_publish": true,
          "levels_completed": 6,
          "actions": 490,
          "resets": 3,
          "state": "WIN",
          "completed": true,
          "number_of_levels": 0,
          "level_scores": [
            115,
            115,
            115,
            4.78515625,
            4.1780388434001825,
            115
          ],
          "level_actions": [
            4,
            7,
            14,
            128,
            318,
            19
          ],
          "level_baseline_actions": [
            43,
            12,
            23,
            28,
            65,
            37
          ]
        },
        {
          "session_id": "09260e6f-af62-40db-965e-23ead9b77691",
          "game_id": "ft09",
          "game_version": "0d8bbf25",
          "published_at": "2026-09-25T05:00:28.80926Z",
          "score": 28.57142857142857,
          "model": "gpt-6-sol",
          "runner": "benchmark_agent",
          "config": "openai-gpt-6-sol-high-provider-adapter",
          "verified_publish": true,
          "levels_completed": 3,
          "actions": 165,
          "resets": 1,
          "state": "NOT_FINISHED",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            115,
            115,
            115,
            0,
            0,
            0
          ],
          "level_actions": [
            4,
            7,
            14,
            140,
            0,
            0
          ],
          "level_baseline_actions": [
            43,
            12,
            23,
            28,
            65,
            37
          ]
        },
        {
          "session_id": "ec226948-ab97-4d64-a9df-30bc40709fe2",
          "game_id": "ft09",
          "game_version": "0d8bbf25",
          "published_at": "2026-09-25T04:49:04.702093Z",
          "score": 4.761904761904762,
          "model": "gpt-6-sol",
          "runner": "benchmark_agent",
          "config": "openai-gpt-6-sol-high",
          "verified_publish": true,
          "levels_completed": 1,
          "actions": 74,
          "resets": 1,
          "state": "NOT_FINISHED",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            115,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            14,
            60,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            43,
            12,
            23,
            28,
            65,
            37
          ]
        },
        {
          "session_id": "6d8a13ec-3a47-4c49-8bed-465699a814c2",
          "game_id": "ft09",
          "game_version": "0d8bbf25",
          "published_at": "2026-09-21T08:52:55.774024Z",
          "score": 84.59745812407792,
          "model": "gemini-3.8-flash",
          "runner": "benchmark_agent",
          "config": "google-gemini-3-8-flash-low-provider-adapter",
          "verified_publish": true,
          "levels_completed": 6,
          "actions": 209,
          "resets": 1,
          "state": "WIN",
          "completed": true,
          "number_of_levels": 0,
          "level_scores": [
            115,
            115,
            115,
            5.727226240046753,
            115,
            81.43961927424152
          ],
          "level_actions": [
            4,
            7,
            14,
            117,
            26,
            41
          ],
          "level_baseline_actions": [
            43,
            12,
            23,
            28,
            65,
            37
          ]
        },
        {
          "session_id": "e37f769d-c244-4a6a-aebe-bd06ae9a4502",
          "game_id": "ft09",
          "game_version": "0d8bbf25",
          "published_at": "2026-09-20T03:57:29.963615Z",
          "score": 100,
          "model": "gemini-3.8-flash",
          "runner": "benchmark_agent",
          "config": "google-gemini-3-8-flash-medium-provider-adapter",
          "verified_publish": true,
          "levels_completed": 6,
          "actions": 100,
          "resets": 0,
          "state": "WIN",
          "completed": true,
          "number_of_levels": 0,
          "level_scores": [
            115,
            115,
            115,
            107.54458161865568,
            115,
            115
          ],
          "level_actions": [
            5,
            7,
            14,
            27,
            28,
            19
          ],
          "level_baseline_actions": [
            43,
            12,
            23,
            28,
            65,
            37
          ]
        },
        {
          "session_id": "0b43e4ff-0f36-4b5e-a80e-bfc15252d3b7",
          "game_id": "ft09",
          "game_version": "0d8bbf25",
          "published_at": "2026-09-17T14:25:51.229812Z",
          "score": 100,
          "model": "gemini-3.8-flash",
          "runner": "benchmark_agent",
          "config": "google-gemini-3-8-flash-high-provider-adapter",
          "verified_publish": true,
          "levels_completed": 6,
          "actions": 83,
          "resets": 0,
          "state": "WIN",
          "completed": true,
          "number_of_levels": 0,
          "level_scores": [
            115,
            115,
            115,
            115,
            115,
            115
          ],
          "level_actions": [
            4,
            7,
            14,
            24,
            21,
            13
          ],
          "level_baseline_actions": [
            43,
            12,
            23,
            28,
            65,
            37
          ]
        },
        {
          "session_id": "039863c7-79c2-45ed-a11b-5e1009c020ae",
          "game_id": "ft09",
          "game_version": "0d8bbf25",
          "published_at": "2026-09-16T19:33:55.504185Z",
          "score": 0,
          "model": "gpt-6-luna",
          "runner": "benchmark_agent",
          "config": "openai-gpt-6-luna-low",
          "verified_publish": true,
          "levels_completed": 0,
          "actions": 215,
          "resets": 0,
          "state": "NOT_FINISHED",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            215,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            43,
            12,
            23,
            28,
            65,
            37
          ]
        },
        {
          "session_id": "e0c6aacd-7060-4cb3-b7ac-6f85d6ee8a31",
          "game_id": "ft09",
          "game_version": "0d8bbf25",
          "published_at": "2026-09-16T19:31:47.704705Z",
          "score": 0,
          "model": "gpt-6-luna",
          "runner": "benchmark_agent",
          "config": "openai-gpt-6-luna-none-provider-adapter",
          "verified_publish": true,
          "levels_completed": 0,
          "actions": 215,
          "resets": 6,
          "state": "NOT_FINISHED",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            215,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            43,
            12,
            23,
            28,
            65,
            37
          ]
        },
        {
          "session_id": "4abd4ae6-dd65-4239-b1bf-825dd5e37d4d",
          "game_id": "ft09",
          "game_version": "0d8bbf25",
          "published_at": "2026-09-16T19:25:29.02671Z",
          "score": 0,
          "model": "gpt-6-luna",
          "runner": "benchmark_agent",
          "config": "openai-gpt-6-luna-none",
          "verified_publish": true,
          "levels_completed": 0,
          "actions": 215,
          "resets": 4,
          "state": "GAME_OVER",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            215,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            43,
            12,
            23,
            28,
            65,
            37
          ]
        },
        {
          "session_id": "d9ee600f-3a07-48f2-8a2f-9856efb41b9c",
          "game_id": "ft09",
          "game_version": "0d8bbf25",
          "published_at": "2026-09-16T17:52:00.61444Z",
          "score": 0,
          "model": "gpt-6-luna",
          "runner": "benchmark_agent",
          "config": "openai-gpt-6-luna-low-provider-adapter",
          "verified_publish": true,
          "levels_completed": 0,
          "actions": 215,
          "resets": 2,
          "state": "GAME_OVER",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            215,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            43,
            12,
            23,
            28,
            65,
            37
          ]
        },
        {
          "session_id": "2d878c36-64e2-44f7-a451-b945124023fa",
          "game_id": "ft09",
          "game_version": "0d8bbf25",
          "published_at": "2026-09-16T16:43:43.47176Z",
          "score": 0,
          "model": "gpt-6-luna",
          "runner": "benchmark_agent",
          "config": "openai-gpt-6-luna-high",
          "verified_publish": true,
          "levels_completed": 0,
          "actions": 215,
          "resets": 0,
          "state": "NOT_FINISHED",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            215,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            43,
            12,
            23,
            28,
            65,
            37
          ]
        },
        {
          "session_id": "f89d3743-7580-4834-a693-deb3af0ad5a1",
          "game_id": "ft09",
          "game_version": "0d8bbf25",
          "published_at": "2026-09-16T09:57:22.285748Z",
          "score": 0,
          "model": "gpt-6-luna",
          "runner": "benchmark_agent",
          "config": "openai-gpt-6-luna-medium-provider-adapter",
          "verified_publish": true,
          "levels_completed": 0,
          "actions": 215,
          "resets": 4,
          "state": "NOT_FINISHED",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            215,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            43,
            12,
            23,
            28,
            65,
            37
          ]
        },
        {
          "session_id": "80965194-27bb-4135-855a-e75872a86c3c",
          "game_id": "ft09",
          "game_version": "0d8bbf25",
          "published_at": "2026-09-16T09:34:32.466149Z",
          "score": 0,
          "model": "gpt-6-luna",
          "runner": "benchmark_agent",
          "config": "openai-gpt-6-luna-medium",
          "verified_publish": true,
          "levels_completed": 0,
          "actions": 215,
          "resets": 0,
          "state": "NOT_FINISHED",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            215,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            43,
            12,
            23,
            28,
            65,
            37
          ]
        },
        {
          "session_id": "7a901fb5-6842-4ee4-b685-ff6981c57efc",
          "game_id": "ft09",
          "game_version": "0d8bbf25",
          "published_at": "2026-09-16T08:49:38.789033Z",
          "score": 0,
          "model": "gpt-6-luna",
          "runner": "benchmark_agent",
          "config": "openai-gpt-6-luna-xhigh",
          "verified_publish": true,
          "levels_completed": 0,
          "actions": 215,
          "resets": 0,
          "state": "NOT_FINISHED",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            215,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            43,
            12,
            23,
            28,
            65,
            37
          ]
        },
        {
          "session_id": "1342fa4b-0fe2-4601-a72d-d55aef89a41a",
          "game_id": "ft09",
          "game_version": "0d8bbf25",
          "published_at": "2026-09-16T08:49:01.42443Z",
          "score": 0.6895420083610231,
          "model": "gpt-6-luna",
          "runner": "benchmark_agent",
          "config": "openai-gpt-6-luna-xhigh-provider-adapter",
          "verified_publish": true,
          "levels_completed": 1,
          "actions": 173,
          "resets": 3,
          "state": "GAME_OVER",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            14.480382175581486,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            113,
            60,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            43,
            12,
            23,
            28,
            65,
            37
          ]
        },
        {
          "session_id": "095d3015-3fa2-48b3-8492-72d84c366943",
          "game_id": "ft09",
          "game_version": "0d8bbf25",
          "published_at": "2026-09-16T08:48:34.015461Z",
          "score": 4.761904761904762,
          "model": "gpt-6-luna",
          "runner": "benchmark_agent",
          "config": "openai-gpt-6-luna-high-provider-adapter",
          "verified_publish": true,
          "levels_completed": 1,
          "actions": 76,
          "resets": 1,
          "state": "GAME_OVER",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            115,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            16,
            60,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            43,
            12,
            23,
            28,
            65,
            37
          ]
        },
        {
          "session_id": "aa8d67e6-9034-44dd-bd07-c066f9ebb063",
          "game_id": "ft09",
          "game_version": "0d8bbf25",
          "published_at": "2026-09-16T07:18:42.757382Z",
          "score": 64.00058788947679,
          "model": "gpt-6-1-sol",
          "runner": "benchmark_agent",
          "config": "openai-gpt-6-1-sol-low",
          "verified_publish": true,
          "levels_completed": 5,
          "actions": 296,
          "resets": 2,
          "state": "GAME_OVER",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            115,
            115,
            115,
            19.753086419753085,
            115,
            0
          ],
          "level_actions": [
            4,
            7,
            14,
            63,
            23,
            185
          ],
          "level_baseline_actions": [
            43,
            12,
            23,
            28,
            65,
            37
          ]
        },
        {
          "session_id": "32da5218-242a-4112-96ad-da16beaed9d5",
          "game_id": "ft09",
          "game_version": "0d8bbf25",
          "published_at": "2026-09-16T06:49:10.468727Z",
          "score": 0,
          "model": "gpt-6-luna",
          "runner": "benchmark_agent",
          "config": "openai-gpt-6-luna-max",
          "verified_publish": true,
          "levels_completed": 0,
          "actions": 215,
          "resets": 0,
          "state": "NOT_FINISHED",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            215,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            43,
            12,
            23,
            28,
            65,
            37
          ]
        },
        {
          "session_id": "036f0606-4348-4826-b681-ba3dfea2c013",
          "game_id": "ft09",
          "game_version": "0d8bbf25",
          "published_at": "2026-09-16T05:44:48.067729Z",
          "score": 96.52345970527789,
          "model": "gpt-6-1-sol",
          "runner": "benchmark_agent",
          "config": "openai-gpt-6-1-sol-medium",
          "verified_publish": true,
          "levels_completed": 6,
          "actions": 149,
          "resets": 0,
          "state": "WIN",
          "completed": true,
          "number_of_levels": 0,
          "level_scores": [
            115,
            115,
            115,
            17.99816345270891,
            115,
            115
          ],
          "level_actions": [
            7,
            7,
            14,
            66,
            21,
            34
          ],
          "level_baseline_actions": [
            43,
            12,
            23,
            28,
            65,
            37
          ]
        },
        {
          "session_id": "93035c32-c959-4236-b0ea-b48122f7cee6",
          "game_id": "ft09",
          "game_version": "0d8bbf25",
          "published_at": "2026-09-16T04:35:01.045432Z",
          "score": 100,
          "model": "gpt-6-1-sol",
          "runner": "benchmark_agent",
          "config": "openai-gpt-6-1-sol-high-provider-adapter",
          "verified_publish": true,
          "levels_completed": 6,
          "actions": 75,
          "resets": 0,
          "state": "WIN",
          "completed": true,
          "number_of_levels": 0,
          "level_scores": [
            115,
            115,
            115,
            115,
            115,
            115
          ],
          "level_actions": [
            4,
            7,
            14,
            16,
            21,
            13
          ],
          "level_baseline_actions": [
            43,
            12,
            23,
            28,
            65,
            37
          ]
        },
        {
          "session_id": "82affb5b-b445-4ded-8327-cd718cd54bfe",
          "game_id": "ft09",
          "game_version": "0d8bbf25",
          "published_at": "2026-09-16T03:53:35.004541Z",
          "score": 100,
          "model": "gpt-6-1-sol",
          "runner": "benchmark_agent",
          "config": "openai-gpt-6-1-sol-high",
          "verified_publish": true,
          "levels_completed": 6,
          "actions": 77,
          "resets": 0,
          "state": "WIN",
          "completed": true,
          "number_of_levels": 0,
          "level_scores": [
            115,
            115,
            115,
            115,
            115,
            115
          ],
          "level_actions": [
            4,
            7,
            14,
            16,
            21,
            15
          ],
          "level_baseline_actions": [
            43,
            12,
            23,
            28,
            65,
            37
          ]
        },
        {
          "session_id": "95a24327-7cc1-46a7-937f-3dd66436766f",
          "game_id": "ft09",
          "game_version": "0d8bbf25",
          "published_at": "2026-09-16T03:24:43.232558Z",
          "score": 100,
          "model": "gpt-6-1-sol",
          "runner": "benchmark_agent",
          "config": "openai-gpt-6-1-sol-xhigh-provider-adapter",
          "verified_publish": true,
          "levels_completed": 6,
          "actions": 75,
          "resets": 0,
          "state": "WIN",
          "completed": true,
          "number_of_levels": 0,
          "level_scores": [
            115,
            115,
            115,
            115,
            115,
            115
          ],
          "level_actions": [
            4,
            7,
            14,
            16,
            21,
            13
          ],
          "level_baseline_actions": [
            43,
            12,
            23,
            28,
            65,
            37
          ]
        },
        {
          "session_id": "9e300bac-844c-4f1e-b2c5-2de4c578eeea",
          "game_id": "ft09",
          "game_version": "0d8bbf25",
          "published_at": "2026-09-16T03:21:59.245838Z",
          "score": 100,
          "model": "gpt-6-1-sol",
          "runner": "benchmark_agent",
          "config": "openai-gpt-6-1-sol-xhigh",
          "verified_publish": true,
          "levels_completed": 6,
          "actions": 86,
          "resets": 0,
          "state": "WIN",
          "completed": true,
          "number_of_levels": 0,
          "level_scores": [
            115,
            115,
            115,
            115,
            115,
            115
          ],
          "level_actions": [
            4,
            7,
            14,
            25,
            21,
            15
          ],
          "level_baseline_actions": [
            43,
            12,
            23,
            28,
            65,
            37
          ]
        },
        {
          "session_id": "68301baf-719d-467e-9808-cf65569d227a",
          "game_id": "ft09",
          "game_version": "0d8bbf25",
          "published_at": "2026-09-16T03:14:03.678104Z",
          "score": 100,
          "model": "gpt-6-1-sol",
          "runner": "benchmark_agent",
          "config": "openai-gpt-6-1-sol-max-provider-adapter",
          "verified_publish": true,
          "levels_completed": 6,
          "actions": 75,
          "resets": 0,
          "state": "WIN",
          "completed": true,
          "number_of_levels": 0,
          "level_scores": [
            115,
            115,
            115,
            115,
            115,
            115
          ],
          "level_actions": [
            4,
            7,
            14,
            16,
            21,
            13
          ],
          "level_baseline_actions": [
            43,
            12,
            23,
            28,
            65,
            37
          ]
        },
        {
          "session_id": "7a12a1ae-2ca7-46ba-93a7-977681a05281",
          "game_id": "ft09",
          "game_version": "0d8bbf25",
          "published_at": "2026-09-16T01:51:45.332555Z",
          "score": 0.4557094304001814,
          "model": "gpt-6-luna",
          "runner": "benchmark_agent",
          "config": "openai-gpt-6-luna-max-provider-adapter",
          "verified_publish": true,
          "levels_completed": 1,
          "actions": 199,
          "resets": 6,
          "state": "NOT_FINISHED",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            9.56989803840381,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            139,
            60,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            43,
            12,
            23,
            28,
            65,
            37
          ]
        },
        {
          "session_id": "9cbf777c-3839-4235-9041-6801ed444794",
          "game_id": "ft09",
          "game_version": "0d8bbf25",
          "published_at": "2026-09-15T19:58:05.435277Z",
          "score": 100,
          "model": "gpt-6-1-sol",
          "runner": "benchmark_agent",
          "config": "openai-gpt-6-1-sol-max",
          "verified_publish": true,
          "levels_completed": 6,
          "actions": 78,
          "resets": 0,
          "state": "WIN",
          "completed": true,
          "number_of_levels": 0,
          "level_scores": [
            115,
            115,
            115,
            115,
            115,
            115
          ],
          "level_actions": [
            4,
            7,
            14,
            16,
            21,
            16
          ],
          "level_baseline_actions": [
            43,
            12,
            23,
            28,
            65,
            37
          ]
        },
        {
          "session_id": "bfab48f6-d7c5-403c-b382-2eaa94a71ff0",
          "game_id": "ft09",
          "game_version": "0d8bbf25",
          "published_at": "2026-09-09T19:29:50.041582Z",
          "score": 14.285714285714285,
          "model": "gemini-3.8-flash",
          "runner": "benchmark_agent",
          "config": "google-gemini-3-8-flash-high",
          "verified_publish": true,
          "levels_completed": 2,
          "actions": 126,
          "resets": 2,
          "state": "NOT_FINISHED",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            115,
            115,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            4,
            7,
            115,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            43,
            12,
            23,
            28,
            65,
            37
          ]
        },
        {
          "session_id": "32b0c74c-107f-458b-ab5c-7816f26dd810",
          "game_id": "ft09",
          "game_version": "0d8bbf25",
          "published_at": "2026-09-09T18:17:48.201576Z",
          "score": 14.285714285714285,
          "model": "gemini-3.8-flash",
          "runner": "benchmark_agent",
          "config": "google-gemini-3-8-flash-medium",
          "verified_publish": true,
          "levels_completed": 2,
          "actions": 126,
          "resets": 12,
          "state": "NOT_FINISHED",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            115,
            115,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            4,
            7,
            115,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            43,
            12,
            23,
            28,
            65,
            37
          ]
        },
        {
          "session_id": "8d0644a2-b11f-4ccd-b7ff-39e4987c2b50",
          "game_id": "ft09",
          "game_version": "0d8bbf25",
          "published_at": "2026-09-09T17:35:50.659393Z",
          "score": 14.285714285714285,
          "model": "gemini-3.8-flash",
          "runner": "benchmark_agent",
          "config": "google-gemini-3-8-flash-low",
          "verified_publish": true,
          "levels_completed": 2,
          "actions": 11,
          "resets": 0,
          "state": "NOT_FINISHED",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            115,
            115,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            4,
            7,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            43,
            12,
            23,
            28,
            65,
            37
          ]
        },
        {
          "session_id": "f59e4495-a788-49d6-9f77-81108963b9a2",
          "game_id": "ft09",
          "game_version": "0d8bbf25",
          "published_at": "2026-08-13T11:10:22.90515Z",
          "score": 47.61904761904761,
          "model": "grok-4.6",
          "runner": "benchmark_agent",
          "config": "xai-grok-4-6-xhigh",
          "verified_publish": true,
          "levels_completed": 4,
          "actions": 433,
          "resets": 2,
          "state": "GAME_OVER",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            42.44719926538108,
            115,
            115,
            115,
            0,
            0
          ],
          "level_actions": [
            66,
            7,
            14,
            21,
            325,
            0
          ],
          "level_baseline_actions": [
            43,
            12,
            23,
            28,
            65,
            37
          ]
        },
        {
          "session_id": "b24debbc-0a0e-4a4f-a225-7deff148af55",
          "game_id": "ft09",
          "game_version": "0d8bbf25",
          "published_at": "2026-07-21T04:34:40.591116Z",
          "score": 100,
          "model": "claude-opus-5",
          "runner": "benchmark_agent",
          "config": "anthropic-claude-opus-5-high",
          "verified_publish": true,
          "levels_completed": 6,
          "actions": 84,
          "resets": 0,
          "state": "WIN",
          "completed": true,
          "number_of_levels": 0,
          "level_scores": [
            115,
            115,
            115,
            115,
            115,
            115
          ],
          "level_actions": [
            4,
            7,
            14,
            21,
            25,
            13
          ],
          "level_baseline_actions": [
            43,
            12,
            23,
            28,
            65,
            37
          ]
        },
        {
          "session_id": "38b506e8-ae7a-4e7c-85ac-307c9037d984",
          "game_id": "ft09",
          "game_version": "0d8bbf25",
          "published_at": "2026-06-24T11:05:57.166305Z",
          "score": 0,
          "model": "gpt-5-6-luna",
          "runner": "benchmark_agent",
          "config": "openai-gpt-5-6-luna-low",
          "verified_publish": true,
          "levels_completed": 0,
          "actions": 215,
          "resets": 0,
          "state": "NOT_FINISHED",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            215,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            43,
            12,
            23,
            28,
            65,
            37
          ]
        },
        {
          "session_id": "b55f26c2-a8b1-4352-918f-628037c4d171",
          "game_id": "ft09",
          "game_version": "0d8bbf25",
          "published_at": "2026-06-24T11:00:24.241717Z",
          "score": 0,
          "model": "gpt-5-6-luna",
          "runner": "benchmark_agent",
          "config": "openai-gpt-5-6-luna-medium",
          "verified_publish": true,
          "levels_completed": 0,
          "actions": 215,
          "resets": 0,
          "state": "NOT_FINISHED",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            215,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            43,
            12,
            23,
            28,
            65,
            37
          ]
        },
        {
          "session_id": "55f8102d-0ed7-49fe-920c-f8931a4b153f",
          "game_id": "ft09",
          "game_version": "0d8bbf25",
          "published_at": "2026-06-24T06:00:30.246381Z",
          "score": 0,
          "model": "gpt-5-6-luna",
          "runner": "benchmark_agent",
          "config": "openai-gpt-5-6-luna-xhigh",
          "verified_publish": true,
          "levels_completed": 0,
          "actions": 215,
          "resets": 8,
          "state": "NOT_FINISHED",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            215,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            43,
            12,
            23,
            28,
            65,
            37
          ]
        },
        {
          "session_id": "6ca1b697-865a-4ca4-9ab0-034645e22900",
          "game_id": "ft09",
          "game_version": "0d8bbf25",
          "published_at": "2026-06-24T05:16:43.076901Z",
          "score": 0,
          "model": "gpt-5-6-luna",
          "runner": "benchmark_agent",
          "config": "openai-gpt-5-6-luna-high",
          "verified_publish": true,
          "levels_completed": 0,
          "actions": 215,
          "resets": 1,
          "state": "NOT_FINISHED",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            215,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            43,
            12,
            23,
            28,
            65,
            37
          ]
        },
        {
          "session_id": "98d74622-ec2c-479b-81e6-c0a344115306",
          "game_id": "ft09",
          "game_version": "0d8bbf25",
          "published_at": "2026-06-23T02:04:05.00529Z",
          "score": 0,
          "model": "gpt-5-6-luna",
          "runner": "benchmark_agent",
          "config": "openai-gpt-5-6-luna-max",
          "verified_publish": true,
          "levels_completed": 0,
          "actions": 36,
          "resets": 4,
          "state": "NOT_FINISHED",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            36,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            43,
            12,
            23,
            28,
            65,
            37
          ]
        },
        {
          "session_id": "74427831-9c14-450c-b287-2e79e5d3c67b",
          "game_id": "ft09",
          "game_version": "0d8bbf25",
          "published_at": "2026-06-20T12:13:51.210932Z",
          "score": 0,
          "model": "gpt-5-6-terra",
          "runner": "benchmark_agent",
          "config": "openai-gpt-5-6-terra-low",
          "verified_publish": true,
          "levels_completed": 0,
          "actions": 215,
          "resets": 1,
          "state": "GAME_OVER",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            215,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            43,
            12,
            23,
            28,
            65,
            37
          ]
        },
        {
          "session_id": "04bd7995-b553-40c4-b913-09e1a76b70f0",
          "game_id": "ft09",
          "game_version": "0d8bbf25",
          "published_at": "2026-06-20T09:40:45.501911Z",
          "score": 0,
          "model": "gpt-5-6-terra",
          "runner": "benchmark_agent",
          "config": "openai-gpt-5-6-terra-medium",
          "verified_publish": true,
          "levels_completed": 0,
          "actions": 215,
          "resets": 5,
          "state": "GAME_OVER",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            215,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            43,
            12,
            23,
            28,
            65,
            37
          ]
        },
        {
          "session_id": "a64de3c2-24a1-4087-9ca1-d5c100bdf5bf",
          "game_id": "ft09",
          "game_version": "0d8bbf25",
          "published_at": "2026-06-20T09:02:14.159947Z",
          "score": 4.761904761904762,
          "model": "gpt-5-6-terra",
          "runner": "benchmark_agent",
          "config": "openai-gpt-5-6-terra-high",
          "verified_publish": true,
          "levels_completed": 1,
          "actions": 71,
          "resets": 4,
          "state": "NOT_FINISHED",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            115,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            11,
            60,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            43,
            12,
            23,
            28,
            65,
            37
          ]
        },
        {
          "session_id": "fdbb8f1c-413c-4119-bbbb-d751284a849e",
          "game_id": "ft09",
          "game_version": "0d8bbf25",
          "published_at": "2026-06-20T02:39:34.975405Z",
          "score": 4.761904761904762,
          "model": "gpt-5-6-terra",
          "runner": "benchmark_agent",
          "config": "openai-gpt-5-6-terra-xhigh",
          "verified_publish": true,
          "levels_completed": 1,
          "actions": 67,
          "resets": 7,
          "state": "NOT_FINISHED",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            115,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            7,
            60,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            43,
            12,
            23,
            28,
            65,
            37
          ]
        },
        {
          "session_id": "a2076a55-d761-4fde-8002-f047c0fac89c",
          "game_id": "ft09",
          "game_version": "0d8bbf25",
          "published_at": "2026-06-19T00:03:01.246034Z",
          "score": 28,
          "model": "gpt-5-6-terra",
          "runner": "benchmark_agent",
          "config": "openai-gpt-5-6-terra-max",
          "verified_publish": true,
          "levels_completed": 3,
          "actions": 182,
          "resets": 7,
          "state": "NOT_FINISHED",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            115,
            64.00000000000001,
            115,
            0,
            0,
            0
          ],
          "level_actions": [
            13,
            15,
            14,
            140,
            0,
            0
          ],
          "level_baseline_actions": [
            43,
            12,
            23,
            28,
            65,
            37
          ]
        },
        {
          "session_id": "e7269903-8865-4616-a0ee-ab6f99e328d7",
          "game_id": "ft09",
          "game_version": "0d8bbf25",
          "published_at": "2026-06-13T04:11:19.048548Z",
          "score": 87.08090609050085,
          "model": "gpt-5-6-sol",
          "runner": "benchmark_agent",
          "config": "openai-gpt-5-6-sol-max",
          "verified_publish": true,
          "levels_completed": 6,
          "actions": 152,
          "resets": 1,
          "state": "WIN",
          "completed": true,
          "number_of_levels": 0,
          "level_scores": [
            115,
            115,
            115,
            115,
            115,
            17.283171316752938
          ],
          "level_actions": [
            4,
            7,
            14,
            16,
            22,
            89
          ],
          "level_baseline_actions": [
            43,
            12,
            23,
            28,
            65,
            37
          ]
        },
        {
          "session_id": "6e6658c6-a674-409e-be0b-a9c6cbfc7a6b",
          "game_id": "ft09",
          "game_version": "0d8bbf25",
          "published_at": "2026-06-12T00:31:54.411559Z",
          "score": 47.61904761904761,
          "model": "gpt-5-6-sol",
          "runner": "benchmark_agent",
          "config": "openai-gpt-5-6-sol-xhigh",
          "verified_publish": true,
          "levels_completed": 4,
          "actions": 296,
          "resets": 5,
          "state": "NOT_FINISHED",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            115,
            115,
            115,
            115,
            0,
            0
          ],
          "level_actions": [
            4,
            7,
            14,
            16,
            255,
            0
          ],
          "level_baseline_actions": [
            43,
            12,
            23,
            28,
            65,
            37
          ]
        },
        {
          "session_id": "294277b3-75b5-4bc9-be6c-a506b08c3f6a",
          "game_id": "ft09",
          "game_version": "0d8bbf25",
          "published_at": "2026-06-10T17:33:00.326849Z",
          "score": 47.61904761904761,
          "model": "gpt-5-6-sol",
          "runner": "benchmark_agent",
          "config": "openai-gpt-5-6-sol-high",
          "verified_publish": true,
          "levels_completed": 4,
          "actions": 374,
          "resets": 33,
          "state": "NOT_FINISHED",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            115,
            115,
            115,
            115,
            0,
            0
          ],
          "level_actions": [
            4,
            7,
            21,
            17,
            325,
            0
          ],
          "level_baseline_actions": [
            43,
            12,
            23,
            28,
            65,
            37
          ]
        },
        {
          "session_id": "faf2d8ba-6a89-4108-9b38-2562dc727d8a",
          "game_id": "ft09",
          "game_version": "0d8bbf25",
          "published_at": "2026-06-10T09:21:31.7604Z",
          "score": 14.285714285714285,
          "model": "gpt-5-6-sol",
          "runner": "benchmark_agent",
          "config": "openai-gpt-5-6-sol-medium",
          "verified_publish": true,
          "levels_completed": 2,
          "actions": 128,
          "resets": 5,
          "state": "NOT_FINISHED",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            115,
            115,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            6,
            7,
            115,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            43,
            12,
            23,
            28,
            65,
            37
          ]
        },
        {
          "session_id": "4ab89f58-7ea8-4929-9265-2a429c9fcb84",
          "game_id": "ft09",
          "game_version": "0d8bbf25",
          "published_at": "2026-06-09T22:23:42.571599Z",
          "score": 4.761904761904762,
          "model": "gpt-5-6-sol",
          "runner": "benchmark_agent",
          "config": "openai-gpt-5-6-sol-low",
          "verified_publish": true,
          "levels_completed": 1,
          "actions": 82,
          "resets": 2,
          "state": "NOT_FINISHED",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            115,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            22,
            60,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            43,
            12,
            23,
            28,
            65,
            37
          ]
        },
        {
          "session_id": "9ed5a8bb-4629-4568-a65f-65ae04955654",
          "game_id": "ft09",
          "game_version": "0d8bbf25",
          "published_at": "2026-05-29T16:01:39.072403Z",
          "score": 28.57142857142857,
          "model": "claude-opus-4-8",
          "runner": "benchmark_agent",
          "config": "anthropic-opus-4-8-high",
          "verified_publish": true,
          "levels_completed": 3,
          "actions": 167,
          "resets": 2,
          "state": "GAME_OVER",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            115,
            115,
            115,
            0,
            0,
            0
          ],
          "level_actions": [
            6,
            7,
            14,
            140,
            0,
            0
          ],
          "level_baseline_actions": [
            43,
            12,
            23,
            28,
            65,
            37
          ]
        },
        {
          "session_id": "920a2bf1-6fd2-40cb-b81c-95993204ad18",
          "game_id": "ft09",
          "game_version": "0d8bbf25",
          "published_at": "2026-05-01T16:15:30.404041Z",
          "score": 4.761904761904762,
          "model": "gpt-5.5-2026-04-23",
          "runner": "benchmark_agent",
          "config": "openai-gpt-5-5-2026-04-23-high",
          "verified_publish": true,
          "levels_completed": 1,
          "actions": 66,
          "resets": 10,
          "state": "NOT_FINISHED",
          "completed": false,
          "number_of_levels": 6,
          "level_scores": [
            115,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            6,
            60,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            43,
            12,
            23,
            28,
            65,
            37
          ]
        },
        {
          "session_id": "2db681dc-2f13-4ef1-afd9-15385d1252a3",
          "game_id": "ft09",
          "game_version": "0d8bbf25",
          "published_at": "2026-05-01T05:07:46.56935Z",
          "score": 4.761904761904762,
          "model": "claude-opus-4-7",
          "runner": "benchmark_agent",
          "config": "anthropic-opus-4-7-high",
          "verified_publish": true,
          "levels_completed": 1,
          "actions": 70,
          "resets": 3,
          "state": "NOT_FINISHED",
          "completed": false,
          "number_of_levels": 6,
          "level_scores": [
            115,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            10,
            60,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            43,
            12,
            23,
            28,
            65,
            37
          ]
        },
        {
          "session_id": "bcd96362-a68b-4990-9458-30f7cca1e85f",
          "game_id": "ft09",
          "game_version": "0d8bbf25",
          "published_at": "2026-04-14T17:43:47.273617Z",
          "score": 0,
          "model": "grok-4.20-beta-0309-reasoning",
          "runner": "benchmark_agent",
          "config": "xai-grok-4-20-beta-0309-reasoning",
          "verified_publish": true,
          "levels_completed": 0,
          "actions": 0,
          "resets": 0,
          "state": "NOT_FINISHED",
          "completed": false,
          "number_of_levels": 6,
          "level_scores": [
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            43,
            12,
            23,
            28,
            65,
            37
          ]
        },
        {
          "session_id": "108d150e-4e25-4eea-ba04-d41c525fe79b",
          "game_id": "ft09",
          "game_version": "0d8bbf25",
          "published_at": "2026-04-14T17:43:46.95485Z",
          "score": 0,
          "model": "gpt-5.4-2026-03-05",
          "runner": "benchmark_agent",
          "config": "openai-gpt-5-4-2026-03-05-high",
          "verified_publish": true,
          "levels_completed": 0,
          "actions": 108,
          "resets": 1,
          "state": "NOT_FINISHED",
          "completed": false,
          "number_of_levels": 6,
          "level_scores": [
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            108,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            43,
            12,
            23,
            28,
            65,
            37
          ]
        },
        {
          "session_id": "f3ed005c-ee70-45b1-9a52-156fee0dac8e",
          "game_id": "ft09",
          "game_version": "0d8bbf25",
          "published_at": "2026-04-14T17:43:46.379391Z",
          "score": 4.761904761904762,
          "model": "gemini-3.1-pro-preview",
          "runner": "benchmark_agent",
          "config": "google-gemini-3-1-pro-preview",
          "verified_publish": true,
          "levels_completed": 1,
          "actions": 48,
          "resets": 0,
          "state": "NOT_FINISHED",
          "completed": false,
          "number_of_levels": 6,
          "level_scores": [
            115,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            40,
            8,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            43,
            12,
            23,
            28,
            65,
            37
          ]
        },
        {
          "session_id": "0833b6a4-70e5-4f61-8bea-ebe0cbbd7afa",
          "game_id": "ft09",
          "game_version": "0d8bbf25",
          "published_at": "2026-04-14T17:43:45.728781Z",
          "score": 0,
          "model": "claude-opus-4-6",
          "runner": "benchmark_agent",
          "config": "anthropic-opus-4-6-max-effort",
          "verified_publish": true,
          "levels_completed": 0,
          "actions": 215,
          "resets": 0,
          "state": "NOT_FINISHED",
          "completed": false,
          "number_of_levels": 6,
          "level_scores": [
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            215,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            43,
            12,
            23,
            28,
            65,
            37
          ]
        }
      ]
    },
    {
      "game": "lp85",
      "source": "https://arcprize.org/tasks/lp85",
      "runs": [
        {
          "session_id": "b631cf11-54d8-4b11-95a1-f564125b83af",
          "game_id": "lp85",
          "game_version": "305b61c3",
          "published_at": "2026-09-25T15:40:23.34024Z",
          "score": 36.55092344998188,
          "model": "gpt-6-sol",
          "runner": "benchmark_agent",
          "config": "openai-gpt-6-sol-xhigh",
          "verified_publish": true,
          "levels_completed": 7,
          "actions": 467,
          "resets": 2,
          "state": "GAME_OVER",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            115,
            115,
            115,
            11.588954277953825,
            64.62898885044214,
            5.450498872049539,
            31.94706994328922,
            0
          ],
          "level_actions": [
            13,
            26,
            19,
            47,
            51,
            257,
            46,
            8
          ],
          "level_baseline_actions": [
            17,
            38,
            31,
            16,
            41,
            60,
            26,
            159
          ]
        },
        {
          "session_id": "47120b19-3d7f-4ef3-b243-50d327af9b98",
          "game_id": "lp85",
          "game_version": "305b61c3",
          "published_at": "2026-09-25T15:28:47.359794Z",
          "score": 69.60482717143731,
          "model": "gpt-6-sol",
          "runner": "benchmark_agent",
          "config": "openai-gpt-6-sol-max",
          "verified_publish": true,
          "levels_completed": 7,
          "actions": 229,
          "resets": 0,
          "state": "NOT_FINISHED",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            115,
            115,
            115,
            26.638917793964616,
            115,
            54.86968449931412,
            115,
            0
          ],
          "level_actions": [
            10,
            9,
            20,
            31,
            21,
            81,
            18,
            39
          ],
          "level_baseline_actions": [
            17,
            38,
            31,
            16,
            41,
            60,
            26,
            159
          ]
        },
        {
          "session_id": "38f0386b-9ba6-4a45-8298-faf2826c8078",
          "game_id": "lp85",
          "game_version": "305b61c3",
          "published_at": "2026-09-25T09:39:53.509178Z",
          "score": 50.59589422966508,
          "model": "gpt-6-sol",
          "runner": "benchmark_agent",
          "config": "openai-gpt-6-sol-high-provider-adapter",
          "verified_publish": true,
          "levels_completed": 7,
          "actions": 1178,
          "resets": 12,
          "state": "GAME_OVER",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            115,
            94.93754109138725,
            93.84765625,
            30.439952437574313,
            115,
            7.6450975811760715,
            70.34339229968784,
            0
          ],
          "level_actions": [
            13,
            39,
            32,
            29,
            22,
            217,
            31,
            795
          ],
          "level_baseline_actions": [
            17,
            38,
            31,
            16,
            41,
            60,
            26,
            159
          ]
        },
        {
          "session_id": "f3e520a5-3848-4999-aadb-5b41f9e045f6",
          "game_id": "lp85",
          "game_version": "305b61c3",
          "published_at": "2026-09-25T07:29:56.057383Z",
          "score": 2.7777777777777777,
          "model": "gpt-6-sol",
          "runner": "benchmark_agent",
          "config": "openai-gpt-6-sol-high",
          "verified_publish": true,
          "levels_completed": 1,
          "actions": 203,
          "resets": 5,
          "state": "GAME_OVER",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            115,
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            13,
            190,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            17,
            38,
            31,
            16,
            41,
            60,
            26,
            159
          ]
        },
        {
          "session_id": "365d4a1e-2101-4d51-bc6f-d9ad1c29dce5",
          "game_id": "lp85",
          "game_version": "305b61c3",
          "published_at": "2026-09-25T06:48:09.404847Z",
          "score": 2.7777777777777777,
          "model": "gpt-6-sol",
          "runner": "benchmark_agent",
          "config": "openai-gpt-6-sol-none-provider-adapter",
          "verified_publish": true,
          "levels_completed": 1,
          "actions": 196,
          "resets": 2,
          "state": "GAME_OVER",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            115,
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            6,
            190,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            17,
            38,
            31,
            16,
            41,
            60,
            26,
            159
          ]
        },
        {
          "session_id": "e7d4f71d-5548-4a40-8cad-8b8b6f3aef3c",
          "game_id": "lp85",
          "game_version": "305b61c3",
          "published_at": "2026-09-25T06:43:40.814044Z",
          "score": 36.858072916666664,
          "model": "gpt-6-sol",
          "runner": "benchmark_agent",
          "config": "openai-gpt-6-sol-medium-provider-adapter",
          "verified_publish": true,
          "levels_completed": 5,
          "actions": 437,
          "resets": 3,
          "state": "GAME_OVER",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            112.890625,
            115,
            115,
            16.000000000000004,
            115,
            0,
            0,
            0
          ],
          "level_actions": [
            16,
            31,
            17,
            40,
            33,
            300,
            0,
            0
          ],
          "level_baseline_actions": [
            17,
            38,
            31,
            16,
            41,
            60,
            26,
            159
          ]
        },
        {
          "session_id": "a917f4fa-7563-47f7-8124-8c2d7fb07ffb",
          "game_id": "lp85",
          "game_version": "305b61c3",
          "published_at": "2026-09-25T06:39:57.354868Z",
          "score": 2.006944444444444,
          "model": "gpt-6-sol",
          "runner": "benchmark_agent",
          "config": "openai-gpt-6-sol-medium",
          "verified_publish": true,
          "levels_completed": 1,
          "actions": 210,
          "resets": 2,
          "state": "GAME_OVER",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            72.24999999999999,
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            20,
            190,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            17,
            38,
            31,
            16,
            41,
            60,
            26,
            159
          ]
        },
        {
          "session_id": "4f79a789-8dab-4442-b53e-07c55b900e09",
          "game_id": "lp85",
          "game_version": "305b61c3",
          "published_at": "2026-09-25T06:35:08.179782Z",
          "score": 2.7777777777777777,
          "model": "gpt-6-sol",
          "runner": "benchmark_agent",
          "config": "openai-gpt-6-sol-low-provider-adapter",
          "verified_publish": true,
          "levels_completed": 1,
          "actions": 201,
          "resets": 2,
          "state": "GAME_OVER",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            115,
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            11,
            190,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            17,
            38,
            31,
            16,
            41,
            60,
            26,
            159
          ]
        },
        {
          "session_id": "4884d67f-dbf2-4698-9991-e5262b76fa90",
          "game_id": "lp85",
          "game_version": "305b61c3",
          "published_at": "2026-09-25T06:31:03.252098Z",
          "score": 2.7777777777777777,
          "model": "gpt-6-sol",
          "runner": "benchmark_agent",
          "config": "openai-gpt-6-sol-none",
          "verified_publish": true,
          "levels_completed": 1,
          "actions": 196,
          "resets": 2,
          "state": "GAME_OVER",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            115,
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            6,
            190,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            17,
            38,
            31,
            16,
            41,
            60,
            26,
            159
          ]
        },
        {
          "session_id": "6feceda3-d193-4dcf-b97f-ff5a3e75a0f7",
          "game_id": "lp85",
          "game_version": "305b61c3",
          "published_at": "2026-09-25T06:26:19.955811Z",
          "score": 2.7777777777777777,
          "model": "gpt-6-sol",
          "runner": "benchmark_agent",
          "config": "openai-gpt-6-sol-low",
          "verified_publish": true,
          "levels_completed": 1,
          "actions": 203,
          "resets": 2,
          "state": "GAME_OVER",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            115,
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            13,
            190,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            17,
            38,
            31,
            16,
            41,
            60,
            26,
            159
          ]
        },
        {
          "session_id": "bb5af101-6188-420d-957a-95fcc1f42fae",
          "game_id": "lp85",
          "game_version": "305b61c3",
          "published_at": "2026-09-25T05:36:47.995856Z",
          "score": 41.66666666666667,
          "model": "gpt-6-sol",
          "runner": "benchmark_agent",
          "config": "openai-gpt-6-sol-xhigh-provider-adapter",
          "verified_publish": true,
          "levels_completed": 5,
          "actions": 400,
          "resets": 3,
          "state": "GAME_OVER",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            115,
            115,
            115,
            79.01234567901234,
            115,
            0,
            0,
            0
          ],
          "level_actions": [
            12,
            31,
            26,
            18,
            13,
            300,
            0,
            0
          ],
          "level_baseline_actions": [
            17,
            38,
            31,
            16,
            41,
            60,
            26,
            159
          ]
        },
        {
          "session_id": "ed1b0a6e-1991-4205-a8e4-67843e1d562c",
          "game_id": "lp85",
          "game_version": "305b61c3",
          "published_at": "2026-09-21T10:47:10.675842Z",
          "score": 100,
          "model": "gemini-3.8-flash",
          "runner": "benchmark_agent",
          "config": "google-gemini-3-8-flash-low-provider-adapter",
          "verified_publish": true,
          "levels_completed": 8,
          "actions": 212,
          "resets": 5,
          "state": "WIN",
          "completed": true,
          "number_of_levels": 0,
          "level_scores": [
            115,
            115,
            115,
            79.01234567901234,
            115,
            115,
            115,
            115
          ],
          "level_actions": [
            8,
            20,
            19,
            18,
            22,
            33,
            9,
            83
          ],
          "level_baseline_actions": [
            17,
            38,
            31,
            16,
            41,
            60,
            26,
            159
          ]
        },
        {
          "session_id": "a828c5f2-9bcb-4b56-9969-5fa2cf729255",
          "game_id": "lp85",
          "game_version": "305b61c3",
          "published_at": "2026-09-20T04:43:38.205044Z",
          "score": 100,
          "model": "gemini-3.8-flash",
          "runner": "benchmark_agent",
          "config": "google-gemini-3-8-flash-medium-provider-adapter",
          "verified_publish": true,
          "levels_completed": 8,
          "actions": 209,
          "resets": 3,
          "state": "WIN",
          "completed": true,
          "number_of_levels": 0,
          "level_scores": [
            115,
            115,
            115,
            70.9141274238227,
            72.96006944444444,
            115,
            115,
            115
          ],
          "level_actions": [
            9,
            26,
            17,
            19,
            48,
            27,
            8,
            55
          ],
          "level_baseline_actions": [
            17,
            38,
            31,
            16,
            41,
            60,
            26,
            159
          ]
        },
        {
          "session_id": "5074015e-0033-4539-b4df-41f8a660da9a",
          "game_id": "lp85",
          "game_version": "305b61c3",
          "published_at": "2026-09-17T14:37:22.524509Z",
          "score": 100,
          "model": "gemini-3.8-flash",
          "runner": "benchmark_agent",
          "config": "google-gemini-3-8-flash-high-provider-adapter",
          "verified_publish": true,
          "levels_completed": 8,
          "actions": 108,
          "resets": 0,
          "state": "WIN",
          "completed": true,
          "number_of_levels": 0,
          "level_scores": [
            115,
            115,
            115,
            113.77777777777777,
            115,
            115,
            115,
            115
          ],
          "level_actions": [
            7,
            8,
            18,
            15,
            10,
            20,
            16,
            14
          ],
          "level_baseline_actions": [
            17,
            38,
            31,
            16,
            41,
            60,
            26,
            159
          ]
        },
        {
          "session_id": "9b8846ff-eea6-4e59-b916-3920ef0c469b",
          "game_id": "lp85",
          "game_version": "305b61c3",
          "published_at": "2026-09-16T19:31:59.315127Z",
          "score": 0.9545514599022327,
          "model": "gpt-6-luna",
          "runner": "benchmark_agent",
          "config": "openai-gpt-6-luna-low-provider-adapter",
          "verified_publish": true,
          "levels_completed": 1,
          "actions": 219,
          "resets": 0,
          "state": "NOT_FINISHED",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            34.363852556480374,
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            29,
            190,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            17,
            38,
            31,
            16,
            41,
            60,
            26,
            159
          ]
        },
        {
          "session_id": "5e147a07-c10f-41ad-8ff4-78b9aa0227f1",
          "game_id": "lp85",
          "game_version": "305b61c3",
          "published_at": "2026-09-16T19:24:20.658024Z",
          "score": 0,
          "model": "gpt-6-luna",
          "runner": "benchmark_agent",
          "config": "openai-gpt-6-luna-low",
          "verified_publish": true,
          "levels_completed": 0,
          "actions": 85,
          "resets": 5,
          "state": "NOT_FINISHED",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            0,
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            85,
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            17,
            38,
            31,
            16,
            41,
            60,
            26,
            159
          ]
        },
        {
          "session_id": "7abd450c-7d7c-4730-87ac-adcef50b71e0",
          "game_id": "lp85",
          "game_version": "305b61c3",
          "published_at": "2026-09-16T19:23:47.649163Z",
          "score": 0,
          "model": "gpt-6-luna",
          "runner": "benchmark_agent",
          "config": "openai-gpt-6-luna-none-provider-adapter",
          "verified_publish": true,
          "levels_completed": 0,
          "actions": 85,
          "resets": 0,
          "state": "NOT_FINISHED",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            0,
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            85,
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            17,
            38,
            31,
            16,
            41,
            60,
            26,
            159
          ]
        },
        {
          "session_id": "05658940-c74d-47ee-b3ed-3bbe954abc8a",
          "game_id": "lp85",
          "game_version": "305b61c3",
          "published_at": "2026-09-16T19:21:52.632347Z",
          "score": 0,
          "model": "gpt-6-luna",
          "runner": "benchmark_agent",
          "config": "openai-gpt-6-luna-none",
          "verified_publish": true,
          "levels_completed": 0,
          "actions": 85,
          "resets": 5,
          "state": "NOT_FINISHED",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            0,
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            85,
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            17,
            38,
            31,
            16,
            41,
            60,
            26,
            159
          ]
        },
        {
          "session_id": "d42268a9-1a5e-4898-95a0-7a5640cef7a1",
          "game_id": "lp85",
          "game_version": "305b61c3",
          "published_at": "2026-09-16T17:51:37.513853Z",
          "score": 2.7777777777777777,
          "model": "gpt-6-luna",
          "runner": "benchmark_agent",
          "config": "openai-gpt-6-luna-max-provider-adapter",
          "verified_publish": true,
          "levels_completed": 1,
          "actions": 198,
          "resets": 2,
          "state": "GAME_OVER",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            115,
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            8,
            190,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            17,
            38,
            31,
            16,
            41,
            60,
            26,
            159
          ]
        },
        {
          "session_id": "f5df0c4b-0d61-421c-a442-4d0e991c1472",
          "game_id": "lp85",
          "game_version": "305b61c3",
          "published_at": "2026-09-16T16:58:01.710043Z",
          "score": 2.7777777777777777,
          "model": "gpt-6-luna",
          "runner": "benchmark_agent",
          "config": "openai-gpt-6-luna-xhigh",
          "verified_publish": true,
          "levels_completed": 1,
          "actions": 200,
          "resets": 1,
          "state": "NOT_FINISHED",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            115,
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            10,
            190,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            17,
            38,
            31,
            16,
            41,
            60,
            26,
            159
          ]
        },
        {
          "session_id": "202438c0-5258-4ebb-b767-81db30229921",
          "game_id": "lp85",
          "game_version": "305b61c3",
          "published_at": "2026-09-16T16:22:27.596075Z",
          "score": 2.7777777777777777,
          "model": "gpt-6-luna",
          "runner": "benchmark_agent",
          "config": "openai-gpt-6-luna-high-provider-adapter",
          "verified_publish": true,
          "levels_completed": 1,
          "actions": 202,
          "resets": 1,
          "state": "GAME_OVER",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            115,
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            12,
            190,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            17,
            38,
            31,
            16,
            41,
            60,
            26,
            159
          ]
        },
        {
          "session_id": "cf3c72a4-0767-4d32-af69-4cb363b030dc",
          "game_id": "lp85",
          "game_version": "305b61c3",
          "published_at": "2026-09-16T15:12:48.451182Z",
          "score": 0.8353566886345241,
          "model": "gpt-6-luna",
          "runner": "benchmark_agent",
          "config": "openai-gpt-6-luna-medium-provider-adapter",
          "verified_publish": true,
          "levels_completed": 1,
          "actions": 221,
          "resets": 2,
          "state": "GAME_OVER",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            30.072840790842868,
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            31,
            190,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            17,
            38,
            31,
            16,
            41,
            60,
            26,
            159
          ]
        },
        {
          "session_id": "7ac12ced-9717-4779-8da8-103855d5558b",
          "game_id": "lp85",
          "game_version": "305b61c3",
          "published_at": "2026-09-16T15:08:57.396969Z",
          "score": 0,
          "model": "gpt-6-luna",
          "runner": "benchmark_agent",
          "config": "openai-gpt-6-luna-high",
          "verified_publish": true,
          "levels_completed": 0,
          "actions": 85,
          "resets": 3,
          "state": "GAME_OVER",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            0,
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            85,
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            17,
            38,
            31,
            16,
            41,
            60,
            26,
            159
          ]
        },
        {
          "session_id": "3da2cb1c-6bd1-4d6d-a779-cb2dfaad6932",
          "game_id": "lp85",
          "game_version": "305b61c3",
          "published_at": "2026-09-16T09:05:39.804069Z",
          "score": 0,
          "model": "gpt-6-luna",
          "runner": "benchmark_agent",
          "config": "openai-gpt-6-luna-medium",
          "verified_publish": true,
          "levels_completed": 0,
          "actions": 85,
          "resets": 2,
          "state": "NOT_FINISHED",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            0,
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            85,
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            17,
            38,
            31,
            16,
            41,
            60,
            26,
            159
          ]
        },
        {
          "session_id": "1909a947-5e2b-422a-89a0-b1eaeed47b25",
          "game_id": "lp85",
          "game_version": "305b61c3",
          "published_at": "2026-09-16T09:02:48.485465Z",
          "score": 0.1959906684027778,
          "model": "gpt-6-luna",
          "runner": "benchmark_agent",
          "config": "openai-gpt-6-luna-xhigh-provider-adapter",
          "verified_publish": true,
          "levels_completed": 1,
          "actions": 254,
          "resets": 3,
          "state": "GAME_OVER",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            7.0556640625,
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            64,
            190,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            17,
            38,
            31,
            16,
            41,
            60,
            26,
            159
          ]
        },
        {
          "session_id": "16a42cc4-016e-4049-a42f-d5b708ec7fa6",
          "game_id": "lp85",
          "game_version": "305b61c3",
          "published_at": "2026-09-16T08:46:42.24143Z",
          "score": 0.39643347050754457,
          "model": "gpt-6-luna",
          "runner": "benchmark_agent",
          "config": "openai-gpt-6-luna-max",
          "verified_publish": true,
          "levels_completed": 1,
          "actions": 235,
          "resets": 14,
          "state": "NOT_FINISHED",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            14.271604938271604,
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            45,
            190,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            17,
            38,
            31,
            16,
            41,
            60,
            26,
            159
          ]
        },
        {
          "session_id": "4d32c559-f9f6-46ec-b438-4ed218224c93",
          "game_id": "lp85",
          "game_version": "305b61c3",
          "published_at": "2026-09-16T07:30:57.710447Z",
          "score": 57.47684757703226,
          "model": "gpt-6-1-sol",
          "runner": "benchmark_agent",
          "config": "openai-gpt-6-1-sol-low",
          "verified_publish": true,
          "levels_completed": 6,
          "actions": 324,
          "resets": 1,
          "state": "NOT_FINISHED",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            80.05540166204986,
            115,
            115,
            113.77777777777777,
            115,
            64.00000000000001,
            0,
            0
          ],
          "level_actions": [
            19,
            29,
            26,
            15,
            30,
            75,
            130,
            0
          ],
          "level_baseline_actions": [
            17,
            38,
            31,
            16,
            41,
            60,
            26,
            159
          ]
        },
        {
          "session_id": "85662a3d-8d7e-4eea-aa56-2824ee6ad281",
          "game_id": "lp85",
          "game_version": "305b61c3",
          "published_at": "2026-09-16T06:45:46.115691Z",
          "score": 83.60670219911499,
          "model": "gpt-6-1-sol",
          "runner": "benchmark_agent",
          "config": "openai-gpt-6-1-sol-medium",
          "verified_publish": true,
          "levels_completed": 8,
          "actions": 516,
          "resets": 5,
          "state": "WIN",
          "completed": true,
          "number_of_levels": 0,
          "level_scores": [
            115,
            115,
            115,
            30.439952437574313,
            115,
            107.01545778834722,
            115,
            21.99859033596993
          ],
          "level_actions": [
            10,
            17,
            16,
            29,
            28,
            58,
            19,
            339
          ],
          "level_baseline_actions": [
            17,
            38,
            31,
            16,
            41,
            60,
            26,
            159
          ]
        },
        {
          "session_id": "888698ac-d270-43c6-b244-c44ae4a0d230",
          "game_id": "lp85",
          "game_version": "305b61c3",
          "published_at": "2026-09-16T04:58:07.351852Z",
          "score": 100,
          "model": "gpt-6-1-sol",
          "runner": "benchmark_agent",
          "config": "openai-gpt-6-1-sol-high-provider-adapter",
          "verified_publish": true,
          "levels_completed": 8,
          "actions": 111,
          "resets": 0,
          "state": "WIN",
          "completed": true,
          "number_of_levels": 0,
          "level_scores": [
            115,
            115,
            115,
            113.77777777777777,
            115,
            115,
            115,
            115
          ],
          "level_actions": [
            11,
            8,
            17,
            15,
            13,
            20,
            8,
            19
          ],
          "level_baseline_actions": [
            17,
            38,
            31,
            16,
            41,
            60,
            26,
            159
          ]
        },
        {
          "session_id": "52aea9d1-5fc6-404c-ae8d-e0167cb20fb6",
          "game_id": "lp85",
          "game_version": "305b61c3",
          "published_at": "2026-09-16T04:08:47.755796Z",
          "score": 100,
          "model": "gpt-6-1-sol",
          "runner": "benchmark_agent",
          "config": "openai-gpt-6-1-sol-high",
          "verified_publish": true,
          "levels_completed": 8,
          "actions": 126,
          "resets": 0,
          "state": "WIN",
          "completed": true,
          "number_of_levels": 0,
          "level_scores": [
            115,
            115,
            115,
            100,
            115,
            115,
            115,
            115
          ],
          "level_actions": [
            10,
            10,
            16,
            16,
            18,
            20,
            9,
            27
          ],
          "level_baseline_actions": [
            17,
            38,
            31,
            16,
            41,
            60,
            26,
            159
          ]
        },
        {
          "session_id": "8a3ef9ea-d875-4c29-b37d-1e3e8d299303",
          "game_id": "lp85",
          "game_version": "305b61c3",
          "published_at": "2026-09-16T03:40:05.164782Z",
          "score": 100,
          "model": "gpt-6-1-sol",
          "runner": "benchmark_agent",
          "config": "openai-gpt-6-1-sol-xhigh",
          "verified_publish": true,
          "levels_completed": 8,
          "actions": 116,
          "resets": 0,
          "state": "WIN",
          "completed": true,
          "number_of_levels": 0,
          "level_scores": [
            115,
            115,
            115,
            64.00000000000001,
            115,
            115,
            115,
            115
          ],
          "level_actions": [
            9,
            8,
            17,
            20,
            12,
            20,
            5,
            25
          ],
          "level_baseline_actions": [
            17,
            38,
            31,
            16,
            41,
            60,
            26,
            159
          ]
        },
        {
          "session_id": "5ccdb415-b7d6-492b-89c4-ea0b4672af91",
          "game_id": "lp85",
          "game_version": "305b61c3",
          "published_at": "2026-09-16T02:48:56.26714Z",
          "score": 100,
          "model": "gpt-6-1-sol",
          "runner": "benchmark_agent",
          "config": "openai-gpt-6-1-sol-xhigh-provider-adapter",
          "verified_publish": true,
          "levels_completed": 8,
          "actions": 104,
          "resets": 0,
          "state": "WIN",
          "completed": true,
          "number_of_levels": 0,
          "level_scores": [
            115,
            115,
            115,
            115,
            115,
            115,
            115,
            115
          ],
          "level_actions": [
            9,
            8,
            16,
            14,
            10,
            20,
            5,
            22
          ],
          "level_baseline_actions": [
            17,
            38,
            31,
            16,
            41,
            60,
            26,
            159
          ]
        },
        {
          "session_id": "54ad1d7c-9834-4331-a39a-ad873765e5fe",
          "game_id": "lp85",
          "game_version": "305b61c3",
          "published_at": "2026-09-16T01:10:31.07462Z",
          "score": 100,
          "model": "gpt-6-1-sol",
          "runner": "benchmark_agent",
          "config": "openai-gpt-6-1-sol-max-provider-adapter",
          "verified_publish": true,
          "levels_completed": 8,
          "actions": 105,
          "resets": 0,
          "state": "WIN",
          "completed": true,
          "number_of_levels": 0,
          "level_scores": [
            115,
            115,
            115,
            100,
            115,
            115,
            115,
            115
          ],
          "level_actions": [
            9,
            8,
            17,
            16,
            10,
            20,
            5,
            20
          ],
          "level_baseline_actions": [
            17,
            38,
            31,
            16,
            41,
            60,
            26,
            159
          ]
        },
        {
          "session_id": "31a1038a-bebc-4044-9394-b007d72d1f03",
          "game_id": "lp85",
          "game_version": "305b61c3",
          "published_at": "2026-09-15T20:22:26.522618Z",
          "score": 100,
          "model": "gpt-6-1-sol",
          "runner": "benchmark_agent",
          "config": "openai-gpt-6-1-sol-max",
          "verified_publish": true,
          "levels_completed": 8,
          "actions": 107,
          "resets": 0,
          "state": "WIN",
          "completed": true,
          "number_of_levels": 0,
          "level_scores": [
            115,
            115,
            115,
            115,
            115,
            115,
            115,
            115
          ],
          "level_actions": [
            9,
            8,
            16,
            13,
            12,
            20,
            14,
            15
          ],
          "level_baseline_actions": [
            17,
            38,
            31,
            16,
            41,
            60,
            26,
            159
          ]
        },
        {
          "session_id": "fa809d8b-7c45-42cf-ad6c-ed4cf319fa91",
          "game_id": "lp85",
          "game_version": "305b61c3",
          "published_at": "2026-09-12T07:28:27.678766Z",
          "score": 86.38575967551066,
          "model": "gemini-3.8-flash",
          "runner": "benchmark_agent",
          "config": "google-gemini-3-8-flash-low",
          "verified_publish": true,
          "levels_completed": 8,
          "actions": 559,
          "resets": 47,
          "state": "WIN",
          "completed": true,
          "number_of_levels": 0,
          "level_scores": [
            115,
            115,
            115,
            88.58131487889274,
            115,
            4.193409358291884,
            115,
            82.55020408163266
          ],
          "level_actions": [
            11,
            17,
            20,
            17,
            17,
            293,
            9,
            175
          ],
          "level_baseline_actions": [
            17,
            38,
            31,
            16,
            41,
            60,
            26,
            159
          ]
        },
        {
          "session_id": "b82b1e2c-dafb-4a99-9ec0-dbfd02bc1325",
          "game_id": "lp85",
          "game_version": "305b61c3",
          "published_at": "2026-09-10T21:49:48.437944Z",
          "score": 41.66666666666667,
          "model": "gemini-3.8-flash",
          "runner": "benchmark_agent",
          "config": "google-gemini-3-8-flash-medium",
          "verified_publish": true,
          "levels_completed": 5,
          "actions": 224,
          "resets": 11,
          "state": "NOT_FINISHED",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            115,
            115,
            115,
            64.00000000000001,
            115,
            0,
            0,
            0
          ],
          "level_actions": [
            12,
            32,
            25,
            20,
            37,
            98,
            0,
            0
          ],
          "level_baseline_actions": [
            17,
            38,
            31,
            16,
            41,
            60,
            26,
            159
          ]
        },
        {
          "session_id": "484de130-939a-4773-b771-d0108ed46a56",
          "game_id": "lp85",
          "game_version": "305b61c3",
          "published_at": "2026-09-10T17:44:27.350953Z",
          "score": 77.77777777777779,
          "model": "gemini-3.8-flash",
          "runner": "benchmark_agent",
          "config": "google-gemini-3-8-flash-high",
          "verified_publish": true,
          "levels_completed": 7,
          "actions": 132,
          "resets": 3,
          "state": "NOT_FINISHED",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            115,
            115,
            115,
            30.439952437574313,
            115,
            115,
            115,
            0
          ],
          "level_actions": [
            8,
            8,
            17,
            29,
            22,
            23,
            9,
            16
          ],
          "level_baseline_actions": [
            17,
            38,
            31,
            16,
            41,
            60,
            26,
            159
          ]
        },
        {
          "session_id": "641c0415-8296-4a7c-b991-fde274be3fe6",
          "game_id": "lp85",
          "game_version": "305b61c3",
          "published_at": "2026-08-11T18:45:33.009851Z",
          "score": 25.04361799816345,
          "model": "grok-4.6",
          "runner": "benchmark_agent",
          "config": "xai-grok-4-6-xhigh",
          "verified_publish": true,
          "levels_completed": 4,
          "actions": 271,
          "resets": 2,
          "state": "GAME_OVER",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            115,
            115,
            115,
            52.892561983471076,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            11,
            15,
            18,
            22,
            205,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            17,
            38,
            31,
            16,
            41,
            60,
            26,
            159
          ]
        },
        {
          "session_id": "39e9c7ff-3e03-444b-b06f-eb45df2151a8",
          "game_id": "lp85",
          "game_version": "305b61c3",
          "published_at": "2026-07-21T09:21:54.701065Z",
          "score": 100,
          "model": "claude-opus-5",
          "runner": "benchmark_agent",
          "config": "anthropic-claude-opus-5-high",
          "verified_publish": true,
          "levels_completed": 8,
          "actions": 108,
          "resets": 0,
          "state": "WIN",
          "completed": true,
          "number_of_levels": 0,
          "level_scores": [
            115,
            115,
            115,
            115,
            115,
            115,
            115,
            115
          ],
          "level_actions": [
            7,
            8,
            18,
            14,
            16,
            20,
            11,
            14
          ],
          "level_baseline_actions": [
            17,
            38,
            31,
            16,
            41,
            60,
            26,
            159
          ]
        },
        {
          "session_id": "117f73df-b5f0-4d2e-8131-f01f3b5312bd",
          "game_id": "lp85",
          "game_version": "305b61c3",
          "published_at": "2026-06-24T11:00:21.580957Z",
          "score": 0,
          "model": "gpt-5-6-luna",
          "runner": "benchmark_agent",
          "config": "openai-gpt-5-6-luna-low",
          "verified_publish": true,
          "levels_completed": 0,
          "actions": 85,
          "resets": 0,
          "state": "NOT_FINISHED",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            0,
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            85,
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            17,
            38,
            31,
            16,
            41,
            60,
            26,
            159
          ]
        },
        {
          "session_id": "098c32e5-3be6-4421-b0d9-1932df760361",
          "game_id": "lp85",
          "game_version": "305b61c3",
          "published_at": "2026-06-24T10:53:18.10631Z",
          "score": 0.8353566886345241,
          "model": "gpt-5-6-luna",
          "runner": "benchmark_agent",
          "config": "openai-gpt-5-6-luna-medium",
          "verified_publish": true,
          "levels_completed": 1,
          "actions": 221,
          "resets": 2,
          "state": "GAME_OVER",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            30.072840790842868,
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            31,
            190,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            17,
            38,
            31,
            16,
            41,
            60,
            26,
            159
          ]
        },
        {
          "session_id": "08716743-9645-4d66-9438-2217e5fb65a3",
          "game_id": "lp85",
          "game_version": "305b61c3",
          "published_at": "2026-06-24T04:20:57.694187Z",
          "score": 0,
          "model": "gpt-5-6-luna",
          "runner": "benchmark_agent",
          "config": "openai-gpt-5-6-luna-high",
          "verified_publish": true,
          "levels_completed": 0,
          "actions": 85,
          "resets": 1,
          "state": "NOT_FINISHED",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            0,
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            85,
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            17,
            38,
            31,
            16,
            41,
            60,
            26,
            159
          ]
        },
        {
          "session_id": "3f22d55a-1eec-4334-834e-c33eca3399e0",
          "game_id": "lp85",
          "game_version": "305b61c3",
          "published_at": "2026-06-24T02:31:10.445393Z",
          "score": 0,
          "model": "gpt-5-6-luna",
          "runner": "benchmark_agent",
          "config": "openai-gpt-5-6-luna-xhigh",
          "verified_publish": true,
          "levels_completed": 0,
          "actions": 85,
          "resets": 7,
          "state": "NOT_FINISHED",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            0,
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            85,
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            17,
            38,
            31,
            16,
            41,
            60,
            26,
            159
          ]
        },
        {
          "session_id": "10688ceb-00f6-431e-bdaa-0b6061f89c7c",
          "game_id": "lp85",
          "game_version": "305b61c3",
          "published_at": "2026-06-23T02:38:32.642472Z",
          "score": 0,
          "model": "gpt-5-6-luna",
          "runner": "benchmark_agent",
          "config": "openai-gpt-5-6-luna-max",
          "verified_publish": true,
          "levels_completed": 0,
          "actions": 85,
          "resets": 20,
          "state": "NOT_FINISHED",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            0,
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            85,
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            17,
            38,
            31,
            16,
            41,
            60,
            26,
            159
          ]
        },
        {
          "session_id": "d46c9c9f-6132-464d-b958-f26ad47735a4",
          "game_id": "lp85",
          "game_version": "305b61c3",
          "published_at": "2026-06-20T12:24:59.436626Z",
          "score": 1.820357772738725,
          "model": "gpt-5-6-terra",
          "runner": "benchmark_agent",
          "config": "openai-gpt-5-6-terra-high",
          "verified_publish": true,
          "levels_completed": 1,
          "actions": 211,
          "resets": 29,
          "state": "GAME_OVER",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            65.5328798185941,
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            21,
            190,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            17,
            38,
            31,
            16,
            41,
            60,
            26,
            159
          ]
        },
        {
          "session_id": "88e0af99-c6a0-4b6f-97d2-7f93be8908dc",
          "game_id": "lp85",
          "game_version": "305b61c3",
          "published_at": "2026-06-20T12:14:32.056934Z",
          "score": 2.7777777777777777,
          "model": "gpt-5-6-terra",
          "runner": "benchmark_agent",
          "config": "openai-gpt-5-6-terra-low",
          "verified_publish": true,
          "levels_completed": 1,
          "actions": 206,
          "resets": 3,
          "state": "NOT_FINISHED",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            112.890625,
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            16,
            190,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            17,
            38,
            31,
            16,
            41,
            60,
            26,
            159
          ]
        },
        {
          "session_id": "fe3106ac-2d63-4017-820f-0beffc692e59",
          "game_id": "lp85",
          "game_version": "305b61c3",
          "published_at": "2026-06-20T09:43:34.397803Z",
          "score": 1.820357772738725,
          "model": "gpt-5-6-terra",
          "runner": "benchmark_agent",
          "config": "openai-gpt-5-6-terra-medium",
          "verified_publish": true,
          "levels_completed": 1,
          "actions": 211,
          "resets": 5,
          "state": "GAME_OVER",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            65.5328798185941,
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            21,
            190,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            17,
            38,
            31,
            16,
            41,
            60,
            26,
            159
          ]
        },
        {
          "session_id": "70408a76-fe64-414f-8567-163ec74417f1",
          "game_id": "lp85",
          "game_version": "305b61c3",
          "published_at": "2026-06-20T08:35:52.884343Z",
          "score": 1.5175383322831335,
          "model": "gpt-5-6-terra",
          "runner": "benchmark_agent",
          "config": "openai-gpt-5-6-terra-xhigh",
          "verified_publish": true,
          "levels_completed": 1,
          "actions": 213,
          "resets": 48,
          "state": "NOT_FINISHED",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            54.6313799621928,
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            23,
            190,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            17,
            38,
            31,
            16,
            41,
            60,
            26,
            159
          ]
        },
        {
          "session_id": "cbf0933c-3a02-4eb7-adab-8d225866257f",
          "game_id": "lp85",
          "game_version": "305b61c3",
          "published_at": "2026-06-19T03:24:28.437585Z",
          "score": 2.7777777777777777,
          "model": "gpt-5-6-terra",
          "runner": "benchmark_agent",
          "config": "openai-gpt-5-6-terra-max",
          "verified_publish": true,
          "levels_completed": 1,
          "actions": 198,
          "resets": 57,
          "state": "NOT_FINISHED",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            115,
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            8,
            190,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            17,
            38,
            31,
            16,
            41,
            60,
            26,
            159
          ]
        },
        {
          "session_id": "637bd72c-e588-4727-97b3-3cee5bb12aba",
          "game_id": "lp85",
          "game_version": "305b61c3",
          "published_at": "2026-06-16T01:42:23.930769Z",
          "score": 37.916666666666664,
          "model": "gpt-5-6-sol",
          "runner": "benchmark_agent",
          "config": "openai-gpt-5-6-sol-xhigh",
          "verified_publish": true,
          "levels_completed": 5,
          "actions": 382,
          "resets": 9,
          "state": "GAME_OVER",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            115,
            115,
            115,
            25,
            115,
            0,
            0,
            0
          ],
          "level_actions": [
            8,
            8,
            16,
            32,
            18,
            300,
            0,
            0
          ],
          "level_baseline_actions": [
            17,
            38,
            31,
            16,
            41,
            60,
            26,
            159
          ]
        },
        {
          "session_id": "4587299f-6aa2-4d20-b135-7608ff20d581",
          "game_id": "lp85",
          "game_version": "305b61c3",
          "published_at": "2026-06-13T17:29:19.333622Z",
          "score": 39.34664694280079,
          "model": "gpt-5-6-sol",
          "runner": "benchmark_agent",
          "config": "openai-gpt-5-6-sol-max",
          "verified_publish": true,
          "levels_completed": 5,
          "actions": 373,
          "resets": 8,
          "state": "GAME_OVER",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            115,
            115,
            115,
            37.8698224852071,
            115,
            0,
            0,
            0
          ],
          "level_actions": [
            7,
            8,
            18,
            26,
            14,
            300,
            0,
            0
          ],
          "level_baseline_actions": [
            17,
            38,
            31,
            16,
            41,
            60,
            26,
            159
          ]
        },
        {
          "session_id": "b54b6dbc-4e70-4a71-9b43-a25309113d4f",
          "game_id": "lp85",
          "game_version": "305b61c3",
          "published_at": "2026-06-12T04:54:32.989436Z",
          "score": 16.666666666666664,
          "model": "gpt-5-6-sol",
          "runner": "benchmark_agent",
          "config": "openai-gpt-5-6-sol-high",
          "verified_publish": true,
          "levels_completed": 3,
          "actions": 135,
          "resets": 0,
          "state": "NOT_FINISHED",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            115,
            115,
            115,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            13,
            24,
            18,
            80,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            17,
            38,
            31,
            16,
            41,
            60,
            26,
            159
          ]
        },
        {
          "session_id": "a2f3d1df-07fc-463b-a04e-99966a02c425",
          "game_id": "lp85",
          "game_version": "305b61c3",
          "published_at": "2026-06-10T10:38:49.772264Z",
          "score": 0.13898507232994767,
          "model": "gpt-5-6-sol",
          "runner": "benchmark_agent",
          "config": "openai-gpt-5-6-sol-medium",
          "verified_publish": true,
          "levels_completed": 1,
          "actions": 266,
          "resets": 8,
          "state": "GAME_OVER",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            5.003462603878116,
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            76,
            190,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            17,
            38,
            31,
            16,
            41,
            60,
            26,
            159
          ]
        },
        {
          "session_id": "fe6542ca-23ef-4cec-a540-3f4aa05f1220",
          "game_id": "lp85",
          "game_version": "305b61c3",
          "published_at": "2026-06-09T23:03:26.763411Z",
          "score": 2.7777777777777777,
          "model": "gpt-5-6-sol",
          "runner": "benchmark_agent",
          "config": "openai-gpt-5-6-sol-low",
          "verified_publish": true,
          "levels_completed": 1,
          "actions": 200,
          "resets": 4,
          "state": "NOT_FINISHED",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            115,
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            10,
            190,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            17,
            38,
            31,
            16,
            41,
            60,
            26,
            159
          ]
        },
        {
          "session_id": "57ee8a2d-b4ea-4685-b5ca-e013f4fccedd",
          "game_id": "lp85",
          "game_version": "305b61c3",
          "published_at": "2026-05-29T20:13:55.94322Z",
          "score": 20.944444444444443,
          "model": "claude-opus-4-8",
          "runner": "benchmark_agent",
          "config": "anthropic-opus-4-8-high",
          "verified_publish": true,
          "levels_completed": 4,
          "actions": 212,
          "resets": 2,
          "state": "GAME_OVER",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            115,
            115,
            115,
            16.000000000000004,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            13,
            28,
            23,
            40,
            108,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            17,
            38,
            31,
            16,
            41,
            60,
            26,
            159
          ]
        },
        {
          "session_id": "ca89df82-4f08-40f9-8bc0-344c88ef91b3",
          "game_id": "lp85",
          "game_version": "305b61c3",
          "published_at": "2026-05-01T05:08:21.389923Z",
          "score": 0,
          "model": "claude-opus-4-7",
          "runner": "benchmark_agent",
          "config": "anthropic-opus-4-7-high",
          "verified_publish": true,
          "levels_completed": 0,
          "actions": 85,
          "resets": 6,
          "state": "GAME_OVER",
          "completed": false,
          "number_of_levels": 8,
          "level_scores": [
            0,
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            85,
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            17,
            38,
            31,
            16,
            41,
            60,
            26,
            159
          ]
        },
        {
          "session_id": "a306029a-1b18-4696-96a4-66e3708e3ebf",
          "game_id": "lp85",
          "game_version": "305b61c3",
          "published_at": "2026-04-30T23:22:37.724635Z",
          "score": 0,
          "model": "gpt-5.5-2026-04-23",
          "runner": "benchmark_agent",
          "config": "openai-gpt-5-5-2026-04-23-high",
          "verified_publish": true,
          "levels_completed": 0,
          "actions": 85,
          "resets": 3,
          "state": "GAME_OVER",
          "completed": false,
          "number_of_levels": 8,
          "level_scores": [
            0,
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            85,
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            17,
            38,
            31,
            16,
            41,
            60,
            26,
            159
          ]
        },
        {
          "session_id": "abd608c0-23bf-4547-be29-bab988768d07",
          "game_id": "lp85",
          "game_version": "305b61c3",
          "published_at": "2026-04-14T17:43:41.226035Z",
          "score": 0,
          "model": "grok-4.20-beta-0309-reasoning",
          "runner": "benchmark_agent",
          "config": "xai-grok-4-20-beta-0309-reasoning",
          "verified_publish": true,
          "levels_completed": 0,
          "actions": 10,
          "resets": 0,
          "state": "NOT_FINISHED",
          "completed": false,
          "number_of_levels": 8,
          "level_scores": [
            0,
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            10,
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            17,
            38,
            31,
            16,
            41,
            60,
            26,
            159
          ]
        },
        {
          "session_id": "a4ccfa94-e996-48d4-b457-97ef2f0f2154",
          "game_id": "lp85",
          "game_version": "305b61c3",
          "published_at": "2026-04-14T17:43:40.992446Z",
          "score": 2.7777777777777777,
          "model": "gpt-5.4-2026-03-05",
          "runner": "benchmark_agent",
          "config": "openai-gpt-5-4-2026-03-05-high",
          "verified_publish": true,
          "levels_completed": 1,
          "actions": 59,
          "resets": 2,
          "state": "NOT_FINISHED",
          "completed": false,
          "number_of_levels": 8,
          "level_scores": [
            115,
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            14,
            45,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            17,
            38,
            31,
            16,
            41,
            60,
            26,
            159
          ]
        },
        {
          "session_id": "fe21dfc8-e297-46b7-87c3-0fc2b9e1e7a0",
          "game_id": "lp85",
          "game_version": "305b61c3",
          "published_at": "2026-04-14T17:43:40.573942Z",
          "score": 0,
          "model": "gemini-3.1-pro-preview",
          "runner": "benchmark_agent",
          "config": "google-gemini-3-1-pro-preview",
          "verified_publish": true,
          "levels_completed": 0,
          "actions": 31,
          "resets": 0,
          "state": "NOT_FINISHED",
          "completed": false,
          "number_of_levels": 8,
          "level_scores": [
            0,
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            31,
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            17,
            38,
            31,
            16,
            41,
            60,
            26,
            159
          ]
        },
        {
          "session_id": "4b5331d6-a3b8-4867-9a16-c94c0ff14195",
          "game_id": "lp85",
          "game_version": "305b61c3",
          "published_at": "2026-04-14T17:43:40.142238Z",
          "score": 2.7777777777777777,
          "model": "claude-opus-4-6",
          "runner": "benchmark_agent",
          "config": "anthropic-opus-4-6-max-effort",
          "verified_publish": true,
          "levels_completed": 1,
          "actions": 201,
          "resets": 4,
          "state": "GAME_OVER",
          "completed": false,
          "number_of_levels": 8,
          "level_scores": [
            115,
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            11,
            190,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            17,
            38,
            31,
            16,
            41,
            60,
            26,
            159
          ]
        }
      ]
    },
    {
      "game": "tn36",
      "source": "https://arcprize.org/tasks/tn36",
      "runs": [
        {
          "session_id": "26067761-5679-49d2-948d-9c17d1bdba51",
          "game_id": "tn36",
          "game_version": "ef4dde99",
          "published_at": "2026-09-25T12:33:16.149869Z",
          "score": 3.249495826032896,
          "model": "gpt-6-sol",
          "runner": "benchmark_agent",
          "config": "openai-gpt-6-sol-max",
          "verified_publish": true,
          "levels_completed": 2,
          "actions": 379,
          "resets": 9,
          "state": "GAME_OVER",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            5.299932715697945,
            42.84297520661157,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            139,
            110,
            130,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            32,
            72,
            26,
            40,
            30,
            55,
            62
          ]
        },
        {
          "session_id": "61c383e5-1f70-4a1d-b053-43b9c0f1a7f9",
          "game_id": "tn36",
          "game_version": "ef4dde99",
          "published_at": "2026-09-25T07:00:36.576815Z",
          "score": 16.929065743944637,
          "model": "gpt-6-sol",
          "runner": "benchmark_agent",
          "config": "openai-gpt-6-sol-medium-provider-adapter",
          "verified_publish": true,
          "levels_completed": 3,
          "actions": 494,
          "resets": 7,
          "state": "GAME_OVER",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            115,
            7.006920415224914,
            115,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            11,
            272,
            11,
            200,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            32,
            72,
            26,
            40,
            30,
            55,
            62
          ]
        },
        {
          "session_id": "255b61fb-2610-4a01-a9c5-00c88d18c105",
          "game_id": "tn36",
          "game_version": "ef4dde99",
          "published_at": "2026-09-25T06:49:31.646587Z",
          "score": 5.204325153140043,
          "model": "gpt-6-sol",
          "runner": "benchmark_agent",
          "config": "openai-gpt-6-sol-xhigh",
          "verified_publish": true,
          "levels_completed": 2,
          "actions": 284,
          "resets": 4,
          "state": "GAME_OVER",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            67.3241288625904,
            39.19848771266541,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            39,
            115,
            130,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            32,
            72,
            26,
            40,
            30,
            55,
            62
          ]
        },
        {
          "session_id": "38e15e2a-dd89-4bc8-b9f9-4c970745e1b6",
          "game_id": "tn36",
          "game_version": "ef4dde99",
          "published_at": "2026-09-25T06:44:42.932041Z",
          "score": 3.571428571428571,
          "model": "gpt-6-sol",
          "runner": "benchmark_agent",
          "config": "openai-gpt-6-sol-none",
          "verified_publish": true,
          "levels_completed": 1,
          "actions": 372,
          "resets": 5,
          "state": "GAME_OVER",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            115,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            12,
            360,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            32,
            72,
            26,
            40,
            30,
            55,
            62
          ]
        },
        {
          "session_id": "1706d06c-a9a0-4bb3-be21-7daa28fd7a18",
          "game_id": "tn36",
          "game_version": "ef4dde99",
          "published_at": "2026-09-25T06:42:48.884552Z",
          "score": 3.571428571428571,
          "model": "gpt-6-sol",
          "runner": "benchmark_agent",
          "config": "openai-gpt-6-sol-high",
          "verified_publish": true,
          "levels_completed": 1,
          "actions": 372,
          "resets": 8,
          "state": "GAME_OVER",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            115,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            12,
            360,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            32,
            72,
            26,
            40,
            30,
            55,
            62
          ]
        },
        {
          "session_id": "8c145e2b-bffa-40d2-9a16-8cc34752fa3d",
          "game_id": "tn36",
          "game_version": "ef4dde99",
          "published_at": "2026-09-25T06:40:51.447811Z",
          "score": 0,
          "model": "gpt-6-sol",
          "runner": "benchmark_agent",
          "config": "openai-gpt-6-sol-none-provider-adapter",
          "verified_publish": true,
          "levels_completed": 0,
          "actions": 160,
          "resets": 5,
          "state": "GAME_OVER",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            160,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            32,
            72,
            26,
            40,
            30,
            55,
            62
          ]
        },
        {
          "session_id": "15532f42-7c5d-4d28-a693-b02366b7899b",
          "game_id": "tn36",
          "game_version": "ef4dde99",
          "published_at": "2026-09-25T06:36:11.739158Z",
          "score": 3.571428571428571,
          "model": "gpt-6-sol",
          "runner": "benchmark_agent",
          "config": "openai-gpt-6-sol-medium",
          "verified_publish": true,
          "levels_completed": 1,
          "actions": 372,
          "resets": 8,
          "state": "GAME_OVER",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            115,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            12,
            360,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            32,
            72,
            26,
            40,
            30,
            55,
            62
          ]
        },
        {
          "session_id": "ee80f56d-5b7a-417c-988a-445d1f89c9e0",
          "game_id": "tn36",
          "game_version": "ef4dde99",
          "published_at": "2026-09-25T06:19:50.484006Z",
          "score": 3.571428571428571,
          "model": "gpt-6-sol",
          "runner": "benchmark_agent",
          "config": "openai-gpt-6-sol-low",
          "verified_publish": true,
          "levels_completed": 1,
          "actions": 374,
          "resets": 6,
          "state": "GAME_OVER",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            115,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            14,
            360,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            32,
            72,
            26,
            40,
            30,
            55,
            62
          ]
        },
        {
          "session_id": "5613b921-1b63-47f1-b25c-9e9e33e76c01",
          "game_id": "tn36",
          "game_version": "ef4dde99",
          "published_at": "2026-09-25T06:16:39.949812Z",
          "score": 0,
          "model": "gpt-6-sol",
          "runner": "benchmark_agent",
          "config": "openai-gpt-6-sol-low-provider-adapter",
          "verified_publish": true,
          "levels_completed": 0,
          "actions": 160,
          "resets": 6,
          "state": "NOT_FINISHED",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            160,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            32,
            72,
            26,
            40,
            30,
            55,
            62
          ]
        },
        {
          "session_id": "0aac9141-e2a0-450a-9143-d11cb9ed399a",
          "game_id": "tn36",
          "game_version": "ef4dde99",
          "published_at": "2026-09-25T05:55:05.15948Z",
          "score": 1.687264375582765,
          "model": "gpt-6-sol",
          "runner": "benchmark_agent",
          "config": "openai-gpt-6-sol-high-provider-adapter",
          "verified_publish": true,
          "levels_completed": 2,
          "actions": 403,
          "resets": 13,
          "state": "NOT_FINISHED",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            22.811316551570503,
            12.216042982373457,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            67,
            206,
            130,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            32,
            72,
            26,
            40,
            30,
            55,
            62
          ]
        },
        {
          "session_id": "5486d41e-0282-42f1-84cb-30e4d8c55f8d",
          "game_id": "tn36",
          "game_version": "ef4dde99",
          "published_at": "2026-09-25T04:39:31.504478Z",
          "score": 8.388397502601457,
          "model": "gpt-6-sol",
          "runner": "benchmark_agent",
          "config": "openai-gpt-6-sol-xhigh-provider-adapter",
          "verified_publish": true,
          "levels_completed": 2,
          "actions": 233,
          "resets": 3,
          "state": "GAME_OVER",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            115,
            59.937565036420395,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            10,
            93,
            130,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            32,
            72,
            26,
            40,
            30,
            55,
            62
          ]
        },
        {
          "session_id": "19dd7c5b-0d6a-4e77-83be-57554995493f",
          "game_id": "tn36",
          "game_version": "ef4dde99",
          "published_at": "2026-09-21T17:03:08.730045Z",
          "score": 21.428571428571427,
          "model": "gemini-3.8-flash",
          "runner": "benchmark_agent",
          "config": "google-gemini-3-8-flash-low-provider-adapter",
          "verified_publish": true,
          "levels_completed": 3,
          "actions": 235,
          "resets": 8,
          "state": "GAME_OVER",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            115,
            115,
            115,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            8,
            14,
            13,
            200,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            32,
            72,
            26,
            40,
            30,
            55,
            62
          ]
        },
        {
          "session_id": "88be5dc4-7b58-45e7-8b50-8f891e55e211",
          "game_id": "tn36",
          "game_version": "ef4dde99",
          "published_at": "2026-09-20T12:27:50.99369Z",
          "score": 50.811941964285715,
          "model": "gemini-3.8-flash",
          "runner": "benchmark_agent",
          "config": "google-gemini-3-8-flash-medium-provider-adapter",
          "verified_publish": true,
          "levels_completed": 5,
          "actions": 481,
          "resets": 5,
          "state": "GAME_OVER",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            115,
            31.640625,
            115,
            115,
            87.890625,
            0,
            0
          ],
          "level_actions": [
            10,
            128,
            11,
            25,
            32,
            275,
            0
          ],
          "level_baseline_actions": [
            32,
            72,
            26,
            40,
            30,
            55,
            62
          ]
        },
        {
          "session_id": "716e3de8-abef-4346-92f7-301e08413c12",
          "game_id": "tn36",
          "game_version": "ef4dde99",
          "published_at": "2026-09-18T02:51:26.109928Z",
          "score": 64.3052252118329,
          "model": "gemini-3.8-flash",
          "runner": "benchmark_agent",
          "config": "google-gemini-3-8-flash-high-provider-adapter",
          "verified_publish": true,
          "levels_completed": 6,
          "actions": 567,
          "resets": 4,
          "state": "GAME_OVER",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            115,
            115,
            115,
            115,
            115,
            12.591050988553592,
            0
          ],
          "level_actions": [
            12,
            41,
            13,
            16,
            20,
            155,
            310
          ],
          "level_baseline_actions": [
            32,
            72,
            26,
            40,
            30,
            55,
            62
          ]
        },
        {
          "session_id": "1c04996c-ea05-42b3-8295-facf7c3f7571",
          "game_id": "tn36",
          "game_version": "ef4dde99",
          "published_at": "2026-09-16T20:31:00.943934Z",
          "score": 0.25825456232913335,
          "model": "gpt-6-luna",
          "runner": "benchmark_agent",
          "config": "openai-gpt-6-luna-medium-provider-adapter",
          "verified_publish": true,
          "levels_completed": 1,
          "actions": 479,
          "resets": 11,
          "state": "GAME_OVER",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            7.231127745215733,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            119,
            360,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            32,
            72,
            26,
            40,
            30,
            55,
            62
          ]
        },
        {
          "session_id": "c801b31c-a351-475f-bac5-bb048b490b42",
          "game_id": "tn36",
          "game_version": "ef4dde99",
          "published_at": "2026-09-16T19:27:56.665259Z",
          "score": 0,
          "model": "gpt-6-luna",
          "runner": "benchmark_agent",
          "config": "openai-gpt-6-luna-none-provider-adapter",
          "verified_publish": true,
          "levels_completed": 0,
          "actions": 160,
          "resets": 2,
          "state": "GAME_OVER",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            160,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            32,
            72,
            26,
            40,
            30,
            55,
            62
          ]
        },
        {
          "session_id": "af5f64cd-5a35-4ac3-a068-0246f3aac433",
          "game_id": "tn36",
          "game_version": "ef4dde99",
          "published_at": "2026-09-16T19:27:09.160602Z",
          "score": 0,
          "model": "gpt-6-luna",
          "runner": "benchmark_agent",
          "config": "openai-gpt-6-luna-low-provider-adapter",
          "verified_publish": true,
          "levels_completed": 0,
          "actions": 160,
          "resets": 6,
          "state": "GAME_OVER",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            160,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            32,
            72,
            26,
            40,
            30,
            55,
            62
          ]
        },
        {
          "session_id": "2d375aed-d7ec-4634-a10a-244050d9acc9",
          "game_id": "tn36",
          "game_version": "ef4dde99",
          "published_at": "2026-09-16T19:25:58.40855Z",
          "score": 0,
          "model": "gpt-6-luna",
          "runner": "benchmark_agent",
          "config": "openai-gpt-6-luna-low",
          "verified_publish": true,
          "levels_completed": 0,
          "actions": 160,
          "resets": 2,
          "state": "GAME_OVER",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            160,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            32,
            72,
            26,
            40,
            30,
            55,
            62
          ]
        },
        {
          "session_id": "8c3d5082-252c-4647-ab94-ba7c979f0fe4",
          "game_id": "tn36",
          "game_version": "ef4dde99",
          "published_at": "2026-09-16T19:25:47.669783Z",
          "score": 0,
          "model": "gpt-6-luna",
          "runner": "benchmark_agent",
          "config": "openai-gpt-6-luna-none",
          "verified_publish": true,
          "levels_completed": 0,
          "actions": 160,
          "resets": 13,
          "state": "GAME_OVER",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            160,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            32,
            72,
            26,
            40,
            30,
            55,
            62
          ]
        },
        {
          "session_id": "5e5e319c-3d17-42ef-9e90-1773e05fb69c",
          "game_id": "tn36",
          "game_version": "ef4dde99",
          "published_at": "2026-09-16T17:00:19.602613Z",
          "score": 2.532647407993668,
          "model": "gpt-6-luna",
          "runner": "benchmark_agent",
          "config": "openai-gpt-6-luna-max",
          "verified_publish": true,
          "levels_completed": 1,
          "actions": 301,
          "resets": 8,
          "state": "GAME_OVER",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            70.9141274238227,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            38,
            263,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            32,
            72,
            26,
            40,
            30,
            55,
            62
          ]
        },
        {
          "session_id": "21495b6b-9bf1-4ab5-b7b5-19002b58c7e2",
          "game_id": "tn36",
          "game_version": "ef4dde99",
          "published_at": "2026-09-16T12:21:50.80662Z",
          "score": 3.571428571428571,
          "model": "gpt-6-luna",
          "runner": "benchmark_agent",
          "config": "openai-gpt-6-luna-max-provider-adapter",
          "verified_publish": true,
          "levels_completed": 1,
          "actions": 367,
          "resets": 6,
          "state": "GAME_OVER",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            115,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            7,
            360,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            32,
            72,
            26,
            40,
            30,
            55,
            62
          ]
        },
        {
          "session_id": "ba3f1009-8ee8-4ce4-b884-8fdb2b640c32",
          "game_id": "tn36",
          "game_version": "ef4dde99",
          "published_at": "2026-09-16T09:55:10.868705Z",
          "score": 0,
          "model": "gpt-6-luna",
          "runner": "benchmark_agent",
          "config": "openai-gpt-6-luna-medium",
          "verified_publish": true,
          "levels_completed": 0,
          "actions": 156,
          "resets": 4,
          "state": "GAME_OVER",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            156,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            32,
            72,
            26,
            40,
            30,
            55,
            62
          ]
        },
        {
          "session_id": "e3424347-d726-4f55-b8f4-40f018a5c0c9",
          "game_id": "tn36",
          "game_version": "ef4dde99",
          "published_at": "2026-09-16T09:03:29.494345Z",
          "score": 0,
          "model": "gpt-6-luna",
          "runner": "benchmark_agent",
          "config": "openai-gpt-6-luna-high",
          "verified_publish": true,
          "levels_completed": 0,
          "actions": 160,
          "resets": 2,
          "state": "NOT_FINISHED",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            160,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            32,
            72,
            26,
            40,
            30,
            55,
            62
          ]
        },
        {
          "session_id": "a5163e4c-9756-4ddf-9f78-ed5131843625",
          "game_id": "tn36",
          "game_version": "ef4dde99",
          "published_at": "2026-09-16T08:51:49.117848Z",
          "score": 0,
          "model": "gpt-6-luna",
          "runner": "benchmark_agent",
          "config": "openai-gpt-6-luna-high-provider-adapter",
          "verified_publish": true,
          "levels_completed": 0,
          "actions": 160,
          "resets": 4,
          "state": "NOT_FINISHED",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            160,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            32,
            72,
            26,
            40,
            30,
            55,
            62
          ]
        },
        {
          "session_id": "033c9185-27ce-486e-bf16-3924051fa0ed",
          "game_id": "tn36",
          "game_version": "ef4dde99",
          "published_at": "2026-09-16T08:31:17.966214Z",
          "score": 0,
          "model": "gpt-6-luna",
          "runner": "benchmark_agent",
          "config": "openai-gpt-6-luna-xhigh",
          "verified_publish": true,
          "levels_completed": 0,
          "actions": 160,
          "resets": 3,
          "state": "NOT_FINISHED",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            160,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            32,
            72,
            26,
            40,
            30,
            55,
            62
          ]
        },
        {
          "session_id": "8e59fffd-fd9f-420e-9f9e-8b8944af0308",
          "game_id": "tn36",
          "game_version": "ef4dde99",
          "published_at": "2026-09-16T08:08:23.894042Z",
          "score": 0,
          "model": "gpt-6-luna",
          "runner": "benchmark_agent",
          "config": "openai-gpt-6-luna-xhigh-provider-adapter",
          "verified_publish": true,
          "levels_completed": 0,
          "actions": 160,
          "resets": 8,
          "state": "NOT_FINISHED",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            160,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            32,
            72,
            26,
            40,
            30,
            55,
            62
          ]
        },
        {
          "session_id": "8bd93153-f1e1-4413-95f3-e9cfbbc72750",
          "game_id": "tn36",
          "game_version": "ef4dde99",
          "published_at": "2026-09-16T07:16:47.235204Z",
          "score": 3.571428571428571,
          "model": "gpt-6-1-sol",
          "runner": "benchmark_agent",
          "config": "openai-gpt-6-1-sol-low",
          "verified_publish": true,
          "levels_completed": 1,
          "actions": 370,
          "resets": 5,
          "state": "GAME_OVER",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            115,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            10,
            360,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            32,
            72,
            26,
            40,
            30,
            55,
            62
          ]
        },
        {
          "session_id": "08ea4989-a7a3-4afd-8244-0b6f77462796",
          "game_id": "tn36",
          "game_version": "ef4dde99",
          "published_at": "2026-09-16T06:42:38.936565Z",
          "score": 75,
          "model": "gpt-6-1-sol",
          "runner": "benchmark_agent",
          "config": "openai-gpt-6-1-sol-xhigh",
          "verified_publish": true,
          "levels_completed": 6,
          "actions": 556,
          "resets": 4,
          "state": "GAME_OVER",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            115,
            32.6530612244898,
            115,
            115,
            115,
            93.10557094490612,
            0
          ],
          "level_actions": [
            9,
            126,
            11,
            17,
            26,
            57,
            310
          ],
          "level_baseline_actions": [
            32,
            72,
            26,
            40,
            30,
            55,
            62
          ]
        },
        {
          "session_id": "148b92c0-6eec-4ef1-a7e7-462e7e5c8550",
          "game_id": "tn36",
          "game_version": "ef4dde99",
          "published_at": "2026-09-16T06:05:49.786101Z",
          "score": 16.93554296473929,
          "model": "gpt-6-1-sol",
          "runner": "benchmark_agent",
          "config": "openai-gpt-6-1-sol-medium",
          "verified_publish": true,
          "levels_completed": 3,
          "actions": 454,
          "resets": 6,
          "state": "GAME_OVER",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            106.55567117585846,
            11.31976591842082,
            115,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            31,
            214,
            9,
            200,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            32,
            72,
            26,
            40,
            30,
            55,
            62
          ]
        },
        {
          "session_id": "9689d7a0-2393-4a05-8c49-5282c1a38377",
          "game_id": "tn36",
          "game_version": "ef4dde99",
          "published_at": "2026-09-16T04:58:27.216098Z",
          "score": 27.96077597826763,
          "model": "gpt-6-1-sol",
          "runner": "benchmark_agent",
          "config": "openai-gpt-6-1-sol-high",
          "verified_publish": true,
          "levels_completed": 4,
          "actions": 275,
          "resets": 5,
          "state": "GAME_OVER",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            115,
            115,
            115,
            23.22543184787342,
            0,
            0,
            0
          ],
          "level_actions": [
            9,
            19,
            14,
            83,
            150,
            0,
            0
          ],
          "level_baseline_actions": [
            32,
            72,
            26,
            40,
            30,
            55,
            62
          ]
        },
        {
          "session_id": "6a4a1502-e959-4a2a-bc64-eb75e3333f8a",
          "game_id": "tn36",
          "game_version": "ef4dde99",
          "published_at": "2026-09-16T04:40:11.343438Z",
          "score": 100,
          "model": "gpt-6-1-sol",
          "runner": "benchmark_agent",
          "config": "openai-gpt-6-1-sol-high-provider-adapter",
          "verified_publish": true,
          "levels_completed": 7,
          "actions": 154,
          "resets": 0,
          "state": "WIN",
          "completed": true,
          "number_of_levels": 0,
          "level_scores": [
            115,
            115,
            115,
            115,
            115,
            115,
            115
          ],
          "level_actions": [
            9,
            13,
            10,
            16,
            20,
            39,
            47
          ],
          "level_baseline_actions": [
            32,
            72,
            26,
            40,
            30,
            55,
            62
          ]
        },
        {
          "session_id": "254d4560-ed9b-434a-912d-10601a186e01",
          "game_id": "tn36",
          "game_version": "ef4dde99",
          "published_at": "2026-09-16T03:41:44.738074Z",
          "score": 100,
          "model": "gpt-6-1-sol",
          "runner": "benchmark_agent",
          "config": "openai-gpt-6-1-sol-xhigh-provider-adapter",
          "verified_publish": true,
          "levels_completed": 7,
          "actions": 206,
          "resets": 0,
          "state": "WIN",
          "completed": true,
          "number_of_levels": 0,
          "level_scores": [
            115,
            115,
            115,
            115,
            115,
            115,
            76.25471136679229
          ],
          "level_actions": [
            10,
            39,
            9,
            14,
            23,
            40,
            71
          ],
          "level_baseline_actions": [
            32,
            72,
            26,
            40,
            30,
            55,
            62
          ]
        },
        {
          "session_id": "03081409-26bc-4a24-b680-f266c3fc6948",
          "game_id": "tn36",
          "game_version": "ef4dde99",
          "published_at": "2026-09-16T02:28:44.550543Z",
          "score": 75,
          "model": "gpt-6-1-sol",
          "runner": "benchmark_agent",
          "config": "openai-gpt-6-1-sol-max",
          "verified_publish": true,
          "levels_completed": 6,
          "actions": 420,
          "resets": 2,
          "state": "GAME_OVER",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            115,
            115,
            115,
            115,
            115,
            115,
            0
          ],
          "level_actions": [
            11,
            11,
            11,
            15,
            21,
            41,
            310
          ],
          "level_baseline_actions": [
            32,
            72,
            26,
            40,
            30,
            55,
            62
          ]
        },
        {
          "session_id": "4a3e792b-0aa4-4f50-b216-16ab4497625c",
          "game_id": "tn36",
          "game_version": "ef4dde99",
          "published_at": "2026-09-16T01:45:09.546518Z",
          "score": 100,
          "model": "gpt-6-1-sol",
          "runner": "benchmark_agent",
          "config": "openai-gpt-6-1-sol-max-provider-adapter",
          "verified_publish": true,
          "levels_completed": 7,
          "actions": 211,
          "resets": 0,
          "state": "WIN",
          "completed": true,
          "number_of_levels": 0,
          "level_scores": [
            115,
            115,
            115,
            115,
            115,
            55.24105186267347,
            115
          ],
          "level_actions": [
            9,
            47,
            9,
            13,
            18,
            74,
            41
          ],
          "level_baseline_actions": [
            32,
            72,
            26,
            40,
            30,
            55,
            62
          ]
        },
        {
          "session_id": "7ab68716-cb0a-4722-b5c3-14ff4fae9d0c",
          "game_id": "tn36",
          "game_version": "ef4dde99",
          "published_at": "2026-09-11T05:59:22.711711Z",
          "score": 10.714285714285714,
          "model": "gemini-3.8-flash",
          "runner": "benchmark_agent",
          "config": "google-gemini-3-8-flash-low",
          "verified_publish": true,
          "levels_completed": 2,
          "actions": 46,
          "resets": 3,
          "state": "NOT_FINISHED",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            115,
            115,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            13,
            29,
            4,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            32,
            72,
            26,
            40,
            30,
            55,
            62
          ]
        },
        {
          "session_id": "62306071-cfc8-4c74-9741-f2eb10366fd0",
          "game_id": "tn36",
          "game_version": "ef4dde99",
          "published_at": "2026-09-11T02:34:09.547634Z",
          "score": 10.714285714285714,
          "model": "gemini-3.8-flash",
          "runner": "benchmark_agent",
          "config": "google-gemini-3-8-flash-medium",
          "verified_publish": true,
          "levels_completed": 2,
          "actions": 179,
          "resets": 3,
          "state": "GAME_OVER",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            115,
            115,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            14,
            35,
            130,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            32,
            72,
            26,
            40,
            30,
            55,
            62
          ]
        },
        {
          "session_id": "df10c72d-d68b-44a5-9246-30707d39ac83",
          "game_id": "tn36",
          "game_version": "ef4dde99",
          "published_at": "2026-09-10T21:46:07.363639Z",
          "score": 10.714285714285714,
          "model": "gemini-3.8-flash",
          "runner": "benchmark_agent",
          "config": "google-gemini-3-8-flash-high",
          "verified_publish": true,
          "levels_completed": 2,
          "actions": 168,
          "resets": 2,
          "state": "NOT_FINISHED",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            115,
            115,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            8,
            30,
            130,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            32,
            72,
            26,
            40,
            30,
            55,
            62
          ]
        },
        {
          "session_id": "d80b28a7-ff08-4a78-af50-6c294a010bca",
          "game_id": "tn36",
          "game_version": "ef4dde99",
          "published_at": "2026-08-11T11:14:46.430428Z",
          "score": 3.571428571428571,
          "model": "grok-4.6",
          "runner": "benchmark_agent",
          "config": "xai-grok-4-6-xhigh",
          "verified_publish": true,
          "levels_completed": 1,
          "actions": 369,
          "resets": 6,
          "state": "GAME_OVER",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            115,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            9,
            360,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            32,
            72,
            26,
            40,
            30,
            55,
            62
          ]
        },
        {
          "session_id": "86560d9a-1c46-4c10-a450-52eeb8bd0901",
          "game_id": "tn36",
          "game_version": "ef4dde99",
          "published_at": "2026-07-21T09:17:45.280955Z",
          "score": 5.996355685131195,
          "model": "claude-opus-5",
          "runner": "benchmark_agent",
          "config": "anthropic-claude-opus-5-high",
          "verified_publish": true,
          "levels_completed": 2,
          "actions": 283,
          "resets": 4,
          "state": "GAME_OVER",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            115,
            26.44897959183673,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            13,
            140,
            130,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            32,
            72,
            26,
            40,
            30,
            55,
            62
          ]
        },
        {
          "session_id": "4d8a693e-1d74-4728-9de2-600ebd3dd175",
          "game_id": "tn36",
          "game_version": "ef4dde99",
          "published_at": "2026-06-24T10:54:26.298177Z",
          "score": 0,
          "model": "gpt-5-6-luna",
          "runner": "benchmark_agent",
          "config": "openai-gpt-5-6-luna-low",
          "verified_publish": true,
          "levels_completed": 0,
          "actions": 160,
          "resets": 2,
          "state": "GAME_OVER",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            160,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            32,
            72,
            26,
            40,
            30,
            55,
            62
          ]
        },
        {
          "session_id": "cb680b32-5cc1-46d7-8e48-33145ffed690",
          "game_id": "tn36",
          "game_version": "ef4dde99",
          "published_at": "2026-06-24T10:32:19.234521Z",
          "score": 0,
          "model": "gpt-5-6-luna",
          "runner": "benchmark_agent",
          "config": "openai-gpt-5-6-luna-medium",
          "verified_publish": true,
          "levels_completed": 0,
          "actions": 160,
          "resets": 3,
          "state": "GAME_OVER",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            160,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            32,
            72,
            26,
            40,
            30,
            55,
            62
          ]
        },
        {
          "session_id": "271fac92-2a68-45d1-a2bd-9134af01c89c",
          "game_id": "tn36",
          "game_version": "ef4dde99",
          "published_at": "2026-06-24T03:28:07.807778Z",
          "score": 0,
          "model": "gpt-5-6-luna",
          "runner": "benchmark_agent",
          "config": "openai-gpt-5-6-luna-high",
          "verified_publish": true,
          "levels_completed": 0,
          "actions": 160,
          "resets": 2,
          "state": "GAME_OVER",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            160,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            32,
            72,
            26,
            40,
            30,
            55,
            62
          ]
        },
        {
          "session_id": "ed93015b-86fb-42c6-b69c-4a313211c547",
          "game_id": "tn36",
          "game_version": "ef4dde99",
          "published_at": "2026-06-24T01:51:18.476038Z",
          "score": 0,
          "model": "gpt-5-6-luna",
          "runner": "benchmark_agent",
          "config": "openai-gpt-5-6-luna-xhigh",
          "verified_publish": true,
          "levels_completed": 0,
          "actions": 160,
          "resets": 21,
          "state": "GAME_OVER",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            160,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            32,
            72,
            26,
            40,
            30,
            55,
            62
          ]
        },
        {
          "session_id": "40ac8337-4c5e-481d-b45e-8d82c18130d6",
          "game_id": "tn36",
          "game_version": "ef4dde99",
          "published_at": "2026-06-23T03:34:58.842313Z",
          "score": 0,
          "model": "gpt-5-6-luna",
          "runner": "benchmark_agent",
          "config": "openai-gpt-5-6-luna-max",
          "verified_publish": true,
          "levels_completed": 0,
          "actions": 160,
          "resets": 32,
          "state": "NOT_FINISHED",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            160,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            32,
            72,
            26,
            40,
            30,
            55,
            62
          ]
        },
        {
          "session_id": "192d5537-29f5-47ab-9234-d9d455649f1a",
          "game_id": "tn36",
          "game_version": "ef4dde99",
          "published_at": "2026-06-20T12:03:10.049701Z",
          "score": 0,
          "model": "gpt-5-6-terra",
          "runner": "benchmark_agent",
          "config": "openai-gpt-5-6-terra-low",
          "verified_publish": true,
          "levels_completed": 0,
          "actions": 160,
          "resets": 2,
          "state": "GAME_OVER",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            160,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            32,
            72,
            26,
            40,
            30,
            55,
            62
          ]
        },
        {
          "session_id": "e00e6409-f151-4718-9253-4395cb5409ce",
          "game_id": "tn36",
          "game_version": "ef4dde99",
          "published_at": "2026-06-20T11:13:53.085939Z",
          "score": 0.3886856049678879,
          "model": "gpt-5-6-terra",
          "runner": "benchmark_agent",
          "config": "openai-gpt-5-6-terra-high",
          "verified_publish": true,
          "levels_completed": 1,
          "actions": 457,
          "resets": 12,
          "state": "GAME_OVER",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            10.883196939100861,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            97,
            360,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            32,
            72,
            26,
            40,
            30,
            55,
            62
          ]
        },
        {
          "session_id": "cf7d3b5d-378e-4e86-882e-6c0f1474ee10",
          "game_id": "tn36",
          "game_version": "ef4dde99",
          "published_at": "2026-06-20T09:27:12.779857Z",
          "score": 0,
          "model": "gpt-5-6-terra",
          "runner": "benchmark_agent",
          "config": "openai-gpt-5-6-terra-medium",
          "verified_publish": true,
          "levels_completed": 0,
          "actions": 160,
          "resets": 2,
          "state": "GAME_OVER",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            160,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            32,
            72,
            26,
            40,
            30,
            55,
            62
          ]
        },
        {
          "session_id": "ad3a6ea0-5f6e-447a-8b26-c970c8803514",
          "game_id": "tn36",
          "game_version": "ef4dde99",
          "published_at": "2026-06-20T07:26:15.823312Z",
          "score": 0.48317384821546533,
          "model": "gpt-5-6-terra",
          "runner": "benchmark_agent",
          "config": "openai-gpt-5-6-terra-xhigh",
          "verified_publish": true,
          "levels_completed": 1,
          "actions": 447,
          "resets": 40,
          "state": "GAME_OVER",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            13.52886775003303,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            87,
            360,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            32,
            72,
            26,
            40,
            30,
            55,
            62
          ]
        },
        {
          "session_id": "7edf45e5-9476-4470-a655-c77b41769836",
          "game_id": "tn36",
          "game_version": "ef4dde99",
          "published_at": "2026-06-15T20:34:23.068604Z",
          "score": 3.133708883598327,
          "model": "gpt-5-6-sol",
          "runner": "benchmark_agent",
          "config": "openai-gpt-5-6-sol-high",
          "verified_publish": true,
          "levels_completed": 2,
          "actions": 342,
          "resets": 21,
          "state": "NOT_FINISHED",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            50.567901234567906,
            18.58797375309262,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            45,
            167,
            130,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            32,
            72,
            26,
            40,
            30,
            55,
            62
          ]
        },
        {
          "session_id": "6d392783-a359-4887-ba26-6a303a84fca4",
          "game_id": "tn36",
          "game_version": "ef4dde99",
          "published_at": "2026-06-13T09:47:51.882346Z",
          "score": 3.571428571428571,
          "model": "gpt-5-6-sol",
          "runner": "benchmark_agent",
          "config": "openai-gpt-5-6-sol-max",
          "verified_publish": true,
          "levels_completed": 1,
          "actions": 374,
          "resets": 12,
          "state": "GAME_OVER",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            115,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            14,
            360,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            32,
            72,
            26,
            40,
            30,
            55,
            62
          ]
        },
        {
          "session_id": "ecfdab5f-c920-477f-8c38-ff4be73e38f3",
          "game_id": "tn36",
          "game_version": "ef4dde99",
          "published_at": "2026-06-12T00:31:13.768851Z",
          "score": 2.0810928431590177,
          "model": "gpt-5-6-sol",
          "runner": "benchmark_agent",
          "config": "openai-gpt-5-6-sol-xhigh",
          "verified_publish": true,
          "levels_completed": 2,
          "actions": 385,
          "resets": 19,
          "state": "GAME_OVER",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            36.45425418298326,
            10.908172712734618,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            53,
            218,
            114,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            32,
            72,
            26,
            40,
            30,
            55,
            62
          ]
        },
        {
          "session_id": "e1655c0f-c730-4597-8cd2-0e906db8ab58",
          "game_id": "tn36",
          "game_version": "ef4dde99",
          "published_at": "2026-06-10T10:08:57.486719Z",
          "score": 2.404433173663943,
          "model": "gpt-5-6-sol",
          "runner": "benchmark_agent",
          "config": "openai-gpt-5-6-sol-medium",
          "verified_publish": true,
          "levels_completed": 1,
          "actions": 399,
          "resets": 28,
          "state": "NOT_FINISHED",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            67.3241288625904,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            39,
            360,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            32,
            72,
            26,
            40,
            30,
            55,
            62
          ]
        },
        {
          "session_id": "ce22e402-1ffc-4eeb-a9a9-3c57a3dd7106",
          "game_id": "tn36",
          "game_version": "ef4dde99",
          "published_at": "2026-06-09T22:34:46.540617Z",
          "score": 3.571428571428571,
          "model": "gpt-5-6-sol",
          "runner": "benchmark_agent",
          "config": "openai-gpt-5-6-sol-low",
          "verified_publish": true,
          "levels_completed": 1,
          "actions": 380,
          "resets": 13,
          "state": "GAME_OVER",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            115,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            20,
            360,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            32,
            72,
            26,
            40,
            30,
            55,
            62
          ]
        },
        {
          "session_id": "13e4fb0b-39e6-4d75-b6cb-f57cd4b83bff",
          "game_id": "tn36",
          "game_version": "ef4dde99",
          "published_at": "2026-05-29T18:43:11.418774Z",
          "score": 0,
          "model": "claude-opus-4-8",
          "runner": "benchmark_agent",
          "config": "anthropic-opus-4-8-high",
          "verified_publish": true,
          "levels_completed": 0,
          "actions": 160,
          "resets": 4,
          "state": "NOT_FINISHED",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            160,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            32,
            72,
            26,
            40,
            30,
            55,
            62
          ]
        },
        {
          "session_id": "f976c594-f096-4093-8644-9b5cc49e3044",
          "game_id": "tn36",
          "game_version": "ef4dde99",
          "published_at": "2026-05-01T05:08:14.470085Z",
          "score": 0,
          "model": "claude-opus-4-7",
          "runner": "benchmark_agent",
          "config": "anthropic-opus-4-7-high",
          "verified_publish": true,
          "levels_completed": 0,
          "actions": 160,
          "resets": 7,
          "state": "NOT_FINISHED",
          "completed": false,
          "number_of_levels": 7,
          "level_scores": [
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            160,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            32,
            72,
            26,
            40,
            30,
            55,
            62
          ]
        },
        {
          "session_id": "6a13d411-205d-483c-84cd-37085c993267",
          "game_id": "tn36",
          "game_version": "ef4dde99",
          "published_at": "2026-04-30T23:22:53.723117Z",
          "score": 3.571428571428571,
          "model": "gpt-5.5-2026-04-23",
          "runner": "benchmark_agent",
          "config": "openai-gpt-5-5-2026-04-23-high",
          "verified_publish": true,
          "levels_completed": 1,
          "actions": 73,
          "resets": 0,
          "state": "NOT_FINISHED",
          "completed": false,
          "number_of_levels": 7,
          "level_scores": [
            115,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            20,
            53,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            32,
            72,
            26,
            40,
            30,
            55,
            62
          ]
        },
        {
          "session_id": "f9397747-41a0-4345-bd54-10d09ad940d3",
          "game_id": "tn36",
          "game_version": "ef4dde99",
          "published_at": "2026-04-14T17:44:28.979872Z",
          "score": 0,
          "model": "grok-4.20-beta-0309-reasoning",
          "runner": "benchmark_agent",
          "config": "xai-grok-4-20-beta-0309-reasoning",
          "verified_publish": true,
          "levels_completed": 0,
          "actions": 62,
          "resets": 1,
          "state": "GAME_OVER",
          "completed": false,
          "number_of_levels": 7,
          "level_scores": [
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            62,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            32,
            72,
            26,
            40,
            30,
            55,
            62
          ]
        },
        {
          "session_id": "6001948a-2e57-4529-93c8-46c2f06255a1",
          "game_id": "tn36",
          "game_version": "ef4dde99",
          "published_at": "2026-04-14T17:44:28.594623Z",
          "score": 0.557406318723191,
          "model": "gpt-5.4-2026-03-05",
          "runner": "benchmark_agent",
          "config": "openai-gpt-5-4-2026-03-05-high",
          "verified_publish": true,
          "levels_completed": 1,
          "actions": 441,
          "resets": 6,
          "state": "GAME_OVER",
          "completed": false,
          "number_of_levels": 7,
          "level_scores": [
            15.60737692424935,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            81,
            360,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            32,
            72,
            26,
            40,
            30,
            55,
            62
          ]
        },
        {
          "session_id": "84f97a2f-b572-41be-8621-e1e9fb6125c5",
          "game_id": "tn36",
          "game_version": "ef4dde99",
          "published_at": "2026-04-14T17:44:27.631402Z",
          "score": 0,
          "model": "gemini-3.1-pro-preview",
          "runner": "benchmark_agent",
          "config": "google-gemini-3-1-pro-preview",
          "verified_publish": true,
          "levels_completed": 0,
          "actions": 160,
          "resets": 8,
          "state": "NOT_FINISHED",
          "completed": false,
          "number_of_levels": 7,
          "level_scores": [
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            160,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            32,
            72,
            26,
            40,
            30,
            55,
            62
          ]
        },
        {
          "session_id": "a9e3d96c-4295-4db1-a68b-20862b5bc2d5",
          "game_id": "tn36",
          "game_version": "ef4dde99",
          "published_at": "2026-04-14T17:44:27.128544Z",
          "score": 3.571428571428571,
          "model": "claude-opus-4-6",
          "runner": "benchmark_agent",
          "config": "anthropic-opus-4-6-max-effort",
          "verified_publish": true,
          "levels_completed": 1,
          "actions": 148,
          "resets": 3,
          "state": "GAME_OVER",
          "completed": false,
          "number_of_levels": 7,
          "level_scores": [
            115,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            11,
            137,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            32,
            72,
            26,
            40,
            30,
            55,
            62
          ]
        }
      ]
    },
    {
      "game": "sp80",
      "source": "https://arcprize.org/tasks/sp80",
      "runs": [
        {
          "session_id": "4fa82faf-0a3d-46d5-8879-a6a1a1894306",
          "game_id": "sp80",
          "game_version": "589a99af",
          "published_at": "2026-09-25T07:54:34.691932Z",
          "score": 4.761904761904762,
          "model": "gpt-6-sol",
          "runner": "benchmark_agent",
          "config": "openai-gpt-6-sol-max",
          "verified_publish": true,
          "levels_completed": 1,
          "actions": 310,
          "resets": 12,
          "state": "GAME_OVER",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            115,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            20,
            290,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            39,
            58,
            25,
            148,
            96,
            152
          ]
        },
        {
          "session_id": "23a746ef-a0be-4e6d-96ed-e65f49ec81b9",
          "game_id": "sp80",
          "game_version": "589a99af",
          "published_at": "2026-09-25T06:56:50.769596Z",
          "score": 4.761904761904762,
          "model": "gpt-6-sol",
          "runner": "benchmark_agent",
          "config": "openai-gpt-6-sol-none-provider-adapter",
          "verified_publish": true,
          "levels_completed": 1,
          "actions": 312,
          "resets": 18,
          "state": "GAME_OVER",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            115,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            22,
            290,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            39,
            58,
            25,
            148,
            96,
            152
          ]
        },
        {
          "session_id": "46414514-858c-4b1f-83cc-61a23bfda1f6",
          "game_id": "sp80",
          "game_version": "589a99af",
          "published_at": "2026-09-25T06:27:23.109778Z",
          "score": 0,
          "model": "gpt-6-sol",
          "runner": "benchmark_agent",
          "config": "openai-gpt-6-sol-none",
          "verified_publish": true,
          "levels_completed": 0,
          "actions": 195,
          "resets": 8,
          "state": "GAME_OVER",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            195,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            39,
            58,
            25,
            148,
            96,
            152
          ]
        },
        {
          "session_id": "3fb0ff21-d0cb-43d1-9163-0feb45fb08e3",
          "game_id": "sp80",
          "game_version": "589a99af",
          "published_at": "2026-09-25T06:20:41.699185Z",
          "score": 14.095982142857142,
          "model": "gpt-6-sol",
          "runner": "benchmark_agent",
          "config": "openai-gpt-6-sol-low-provider-adapter",
          "verified_publish": true,
          "levels_completed": 2,
          "actions": 182,
          "resets": 9,
          "state": "GAME_OVER",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            66.015625,
            115,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            48,
            9,
            125,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            39,
            58,
            25,
            148,
            96,
            152
          ]
        },
        {
          "session_id": "f40ecdb6-8c95-4cf7-bc74-3050eee0af15",
          "game_id": "sp80",
          "game_version": "589a99af",
          "published_at": "2026-09-25T05:58:58.680408Z",
          "score": 8.555592171509126,
          "model": "gpt-6-sol",
          "runner": "benchmark_agent",
          "config": "openai-gpt-6-sol-medium-provider-adapter",
          "verified_publish": true,
          "levels_completed": 2,
          "actions": 241,
          "resets": 8,
          "state": "GAME_OVER",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            115,
            32.333717800845825,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            14,
            102,
            125,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            39,
            58,
            25,
            148,
            96,
            152
          ]
        },
        {
          "session_id": "9dd001bf-fef3-434e-ade5-8d635406cbc7",
          "game_id": "sp80",
          "game_version": "589a99af",
          "published_at": "2026-09-25T05:58:25.823003Z",
          "score": 0,
          "model": "gpt-6-sol",
          "runner": "benchmark_agent",
          "config": "openai-gpt-6-sol-low",
          "verified_publish": true,
          "levels_completed": 0,
          "actions": 195,
          "resets": 9,
          "state": "GAME_OVER",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            195,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            39,
            58,
            25,
            148,
            96,
            152
          ]
        },
        {
          "session_id": "eeff6f6f-66fe-4e59-8f33-92778e2a7fa0",
          "game_id": "sp80",
          "game_version": "589a99af",
          "published_at": "2026-09-25T05:50:05.852661Z",
          "score": 0.207122226625215,
          "model": "gpt-6-sol",
          "runner": "benchmark_agent",
          "config": "openai-gpt-6-sol-high",
          "verified_publish": true,
          "levels_completed": 1,
          "actions": 477,
          "resets": 15,
          "state": "GAME_OVER",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            4.349566759129515,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            187,
            290,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            39,
            58,
            25,
            148,
            96,
            152
          ]
        },
        {
          "session_id": "b04fe62c-ed6f-40c3-8b21-f4c71314f02f",
          "game_id": "sp80",
          "game_version": "589a99af",
          "published_at": "2026-09-25T05:33:00.212371Z",
          "score": 4.761904761904762,
          "model": "gpt-6-sol",
          "runner": "benchmark_agent",
          "config": "openai-gpt-6-sol-xhigh",
          "verified_publish": true,
          "levels_completed": 1,
          "actions": 324,
          "resets": 11,
          "state": "GAME_OVER",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            115,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            34,
            290,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            39,
            58,
            25,
            148,
            96,
            152
          ]
        },
        {
          "session_id": "95db759a-4ce8-4c14-b34c-73f34832b8e0",
          "game_id": "sp80",
          "game_version": "589a99af",
          "published_at": "2026-09-25T05:22:42.848317Z",
          "score": 0,
          "model": "gpt-6-sol",
          "runner": "benchmark_agent",
          "config": "openai-gpt-6-sol-medium",
          "verified_publish": true,
          "levels_completed": 0,
          "actions": 195,
          "resets": 6,
          "state": "GAME_OVER",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            195,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            39,
            58,
            25,
            148,
            96,
            152
          ]
        },
        {
          "session_id": "17148e14-687c-4bde-b92d-8f812244275d",
          "game_id": "sp80",
          "game_version": "589a99af",
          "published_at": "2026-09-25T05:16:45.118525Z",
          "score": 7.110787172011662,
          "model": "gpt-6-sol",
          "runner": "benchmark_agent",
          "config": "openai-gpt-6-sol-high-provider-adapter",
          "verified_publish": true,
          "levels_completed": 2,
          "actions": 276,
          "resets": 11,
          "state": "GAME_OVER",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            115,
            17.16326530612245,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            11,
            140,
            125,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            39,
            58,
            25,
            148,
            96,
            152
          ]
        },
        {
          "session_id": "5fd99826-f6ac-4e1b-bcc0-6912db3c2dae",
          "game_id": "sp80",
          "game_version": "589a99af",
          "published_at": "2026-09-25T04:16:06.593323Z",
          "score": 14.285714285714285,
          "model": "gpt-6-sol",
          "runner": "benchmark_agent",
          "config": "openai-gpt-6-sol-xhigh-provider-adapter",
          "verified_publish": true,
          "levels_completed": 2,
          "actions": 146,
          "resets": 5,
          "state": "GAME_OVER",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            115,
            115,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            12,
            9,
            125,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            39,
            58,
            25,
            148,
            96,
            152
          ]
        },
        {
          "session_id": "6b24ca2d-7968-4c80-beed-a9745cb1b6b0",
          "game_id": "sp80",
          "game_version": "589a99af",
          "published_at": "2026-09-21T19:08:55.27679Z",
          "score": 17.060576067904393,
          "model": "gemini-3.8-flash",
          "runner": "benchmark_agent",
          "config": "google-gemini-3-8-flash-low-provider-adapter",
          "verified_publish": true,
          "levels_completed": 5,
          "actions": 1694,
          "resets": 58,
          "state": "GAME_OVER",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            115,
            19.905325443786985,
            33.802055164954034,
            18.72712971512602,
            5.429352436610426,
            0
          ],
          "level_actions": [
            7,
            130,
            43,
            342,
            412,
            760
          ],
          "level_baseline_actions": [
            39,
            58,
            25,
            148,
            96,
            152
          ]
        },
        {
          "session_id": "3061de93-903b-459d-9b99-e2dd111fef85",
          "game_id": "sp80",
          "game_version": "589a99af",
          "published_at": "2026-09-20T09:47:28.919374Z",
          "score": 14.285714285714285,
          "model": "gemini-3.8-flash",
          "runner": "benchmark_agent",
          "config": "google-gemini-3-8-flash-medium-provider-adapter",
          "verified_publish": true,
          "levels_completed": 2,
          "actions": 182,
          "resets": 4,
          "state": "GAME_OVER",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            115,
            115,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            9,
            48,
            125,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            39,
            58,
            25,
            148,
            96,
            152
          ]
        },
        {
          "session_id": "3d544a94-5088-4953-a003-196e31badd86",
          "game_id": "sp80",
          "game_version": "589a99af",
          "published_at": "2026-09-18T01:57:09.275161Z",
          "score": 47.61904761904761,
          "model": "gemini-3.8-flash",
          "runner": "benchmark_agent",
          "config": "google-gemini-3-8-flash-high-provider-adapter",
          "verified_publish": true,
          "levels_completed": 4,
          "actions": 577,
          "resets": 7,
          "state": "GAME_OVER",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            115,
            115,
            115,
            115,
            0,
            0
          ],
          "level_actions": [
            8,
            9,
            10,
            70,
            480,
            0
          ],
          "level_baseline_actions": [
            39,
            58,
            25,
            148,
            96,
            152
          ]
        },
        {
          "session_id": "21c7498d-ba36-4f78-842b-9c4dcb3c4b13",
          "game_id": "sp80",
          "game_version": "589a99af",
          "published_at": "2026-09-16T21:03:01.362534Z",
          "score": 0,
          "model": "gpt-6-luna",
          "runner": "benchmark_agent",
          "config": "openai-gpt-6-luna-none-provider-adapter",
          "verified_publish": true,
          "levels_completed": 0,
          "actions": 195,
          "resets": 9,
          "state": "GAME_OVER",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            195,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            39,
            58,
            25,
            148,
            96,
            152
          ]
        },
        {
          "session_id": "3461f65e-e946-46f1-9203-0147fdbdaa59",
          "game_id": "sp80",
          "game_version": "589a99af",
          "published_at": "2026-09-16T20:26:10.38077Z",
          "score": 2.0119047619047623,
          "model": "gpt-6-luna",
          "runner": "benchmark_agent",
          "config": "openai-gpt-6-luna-low-provider-adapter",
          "verified_publish": true,
          "levels_completed": 1,
          "actions": 191,
          "resets": 14,
          "state": "GAME_OVER",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            42.25000000000001,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            60,
            131,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            39,
            58,
            25,
            148,
            96,
            152
          ]
        },
        {
          "session_id": "116aa685-7d63-42cf-b95a-d1364a387909",
          "game_id": "sp80",
          "game_version": "589a99af",
          "published_at": "2026-09-16T20:06:23.009139Z",
          "score": 4.761904761904762,
          "model": "gpt-6-luna",
          "runner": "benchmark_agent",
          "config": "openai-gpt-6-luna-high-provider-adapter",
          "verified_publish": true,
          "levels_completed": 1,
          "actions": 306,
          "resets": 12,
          "state": "GAME_OVER",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            115,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            16,
            290,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            39,
            58,
            25,
            148,
            96,
            152
          ]
        },
        {
          "session_id": "b60856f9-0d6b-4c7a-a351-f86da50d9bfe",
          "game_id": "sp80",
          "game_version": "589a99af",
          "published_at": "2026-09-16T19:52:02.997938Z",
          "score": 4.761904761904762,
          "model": "gpt-6-luna",
          "runner": "benchmark_agent",
          "config": "openai-gpt-6-luna-medium-provider-adapter",
          "verified_publish": true,
          "levels_completed": 1,
          "actions": 313,
          "resets": 21,
          "state": "GAME_OVER",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            115,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            23,
            290,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            39,
            58,
            25,
            148,
            96,
            152
          ]
        },
        {
          "session_id": "466fddd2-821b-4589-9cbe-c94a04adade0",
          "game_id": "sp80",
          "game_version": "589a99af",
          "published_at": "2026-09-16T19:29:26.285383Z",
          "score": 0,
          "model": "gpt-6-luna",
          "runner": "benchmark_agent",
          "config": "openai-gpt-6-luna-low",
          "verified_publish": true,
          "levels_completed": 0,
          "actions": 195,
          "resets": 6,
          "state": "GAME_OVER",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            195,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            39,
            58,
            25,
            148,
            96,
            152
          ]
        },
        {
          "session_id": "c9f97039-6c47-4757-8f3b-447b1beb1551",
          "game_id": "sp80",
          "game_version": "589a99af",
          "published_at": "2026-09-16T19:25:30.0423Z",
          "score": 0,
          "model": "gpt-6-luna",
          "runner": "benchmark_agent",
          "config": "openai-gpt-6-luna-none",
          "verified_publish": true,
          "levels_completed": 0,
          "actions": 195,
          "resets": 9,
          "state": "GAME_OVER",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            195,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            39,
            58,
            25,
            148,
            96,
            152
          ]
        },
        {
          "session_id": "b9644c86-89d3-4638-bb61-1efb4cf8d362",
          "game_id": "sp80",
          "game_version": "589a99af",
          "published_at": "2026-09-16T14:47:36.801032Z",
          "score": 4.761904761904762,
          "model": "gpt-6-luna",
          "runner": "benchmark_agent",
          "config": "openai-gpt-6-luna-xhigh-provider-adapter",
          "verified_publish": true,
          "levels_completed": 1,
          "actions": 64,
          "resets": 3,
          "state": "GAME_OVER",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            115,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            19,
            45,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            39,
            58,
            25,
            148,
            96,
            152
          ]
        },
        {
          "session_id": "0f7a2ead-8036-4ea2-a40d-97d6304e5161",
          "game_id": "sp80",
          "game_version": "589a99af",
          "published_at": "2026-09-16T09:08:47.911678Z",
          "score": 0.9792938267789537,
          "model": "gpt-6-luna",
          "runner": "benchmark_agent",
          "config": "openai-gpt-6-luna-medium",
          "verified_publish": true,
          "levels_completed": 1,
          "actions": 246,
          "resets": 8,
          "state": "GAME_OVER",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            20.565170362358028,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            86,
            160,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            39,
            58,
            25,
            148,
            96,
            152
          ]
        },
        {
          "session_id": "c47de750-9928-4328-b1a6-e0f276c02ce9",
          "game_id": "sp80",
          "game_version": "589a99af",
          "published_at": "2026-09-16T09:00:41.812421Z",
          "score": 4.761904761904762,
          "model": "gpt-6-luna",
          "runner": "benchmark_agent",
          "config": "openai-gpt-6-luna-high",
          "verified_publish": true,
          "levels_completed": 1,
          "actions": 329,
          "resets": 7,
          "state": "GAME_OVER",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            100,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            39,
            290,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            39,
            58,
            25,
            148,
            96,
            152
          ]
        },
        {
          "session_id": "eec1f0a2-a236-436f-b62c-7c6d9d4eba94",
          "game_id": "sp80",
          "game_version": "589a99af",
          "published_at": "2026-09-16T08:46:39.827177Z",
          "score": 4.761904761904762,
          "model": "gpt-6-luna",
          "runner": "benchmark_agent",
          "config": "openai-gpt-6-luna-max-provider-adapter",
          "verified_publish": true,
          "levels_completed": 1,
          "actions": 318,
          "resets": 17,
          "state": "GAME_OVER",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            115,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            28,
            290,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            39,
            58,
            25,
            148,
            96,
            152
          ]
        },
        {
          "session_id": "869bb0f2-7d65-4780-b625-51630fe40b64",
          "game_id": "sp80",
          "game_version": "589a99af",
          "published_at": "2026-09-16T08:18:20.280972Z",
          "score": 4.761904761904762,
          "model": "gpt-6-luna",
          "runner": "benchmark_agent",
          "config": "openai-gpt-6-luna-xhigh",
          "verified_publish": true,
          "levels_completed": 1,
          "actions": 302,
          "resets": 8,
          "state": "GAME_OVER",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            115,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            12,
            290,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            39,
            58,
            25,
            148,
            96,
            152
          ]
        },
        {
          "session_id": "ae63f738-a7c1-4193-8ff3-a1dbe84a2444",
          "game_id": "sp80",
          "game_version": "589a99af",
          "published_at": "2026-09-16T06:52:56.82892Z",
          "score": 14.285714285714285,
          "model": "gpt-6-1-sol",
          "runner": "benchmark_agent",
          "config": "openai-gpt-6-1-sol-low",
          "verified_publish": true,
          "levels_completed": 2,
          "actions": 145,
          "resets": 4,
          "state": "GAME_OVER",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            115,
            115,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            6,
            14,
            125,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            39,
            58,
            25,
            148,
            96,
            152
          ]
        },
        {
          "session_id": "9f9b3818-03f5-4325-90e4-d92ce33b004a",
          "game_id": "sp80",
          "game_version": "589a99af",
          "published_at": "2026-09-16T06:42:53.731694Z",
          "score": 4.761904761904762,
          "model": "gpt-6-luna",
          "runner": "benchmark_agent",
          "config": "openai-gpt-6-luna-max",
          "verified_publish": true,
          "levels_completed": 1,
          "actions": 319,
          "resets": 8,
          "state": "GAME_OVER",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            115,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            29,
            290,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            39,
            58,
            25,
            148,
            96,
            152
          ]
        },
        {
          "session_id": "bcd06c91-aede-4bc9-90af-a5404ded651d",
          "game_id": "sp80",
          "game_version": "589a99af",
          "published_at": "2026-09-16T06:33:10.389254Z",
          "score": 60.934966733705366,
          "model": "gpt-6-1-sol",
          "runner": "benchmark_agent",
          "config": "openai-gpt-6-1-sol-high",
          "verified_publish": true,
          "levels_completed": 6,
          "actions": 819,
          "resets": 22,
          "state": "WIN",
          "completed": true,
          "number_of_levels": 0,
          "level_scores": [
            115,
            115,
            115,
            115,
            14.284828569657137,
            9.701693093254503
          ],
          "level_actions": [
            9,
            8,
            11,
            49,
            254,
            488
          ],
          "level_baseline_actions": [
            39,
            58,
            25,
            148,
            96,
            152
          ]
        },
        {
          "session_id": "38580d12-70c7-407f-ade1-05d0cd7c4c45",
          "game_id": "sp80",
          "game_version": "589a99af",
          "published_at": "2026-09-16T05:29:49.395265Z",
          "score": 14.285714285714285,
          "model": "gpt-6-1-sol",
          "runner": "benchmark_agent",
          "config": "openai-gpt-6-1-sol-medium",
          "verified_publish": true,
          "levels_completed": 2,
          "actions": 177,
          "resets": 6,
          "state": "GAME_OVER",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            115,
            115,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            7,
            45,
            125,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            39,
            58,
            25,
            148,
            96,
            152
          ]
        },
        {
          "session_id": "ce9c1b6b-343c-48de-9433-e4d5e6005c3f",
          "game_id": "sp80",
          "game_version": "589a99af",
          "published_at": "2026-09-16T05:12:03.401248Z",
          "score": 94.54458010953402,
          "model": "gpt-6-1-sol",
          "runner": "benchmark_agent",
          "config": "openai-gpt-6-1-sol-xhigh",
          "verified_publish": true,
          "levels_completed": 6,
          "actions": 377,
          "resets": 4,
          "state": "WIN",
          "completed": true,
          "number_of_levels": 0,
          "level_scores": [
            115,
            115,
            115,
            115,
            29.08723646004292,
            115
          ],
          "level_actions": [
            7,
            9,
            10,
            56,
            178,
            117
          ],
          "level_baseline_actions": [
            39,
            58,
            25,
            148,
            96,
            152
          ]
        },
        {
          "session_id": "7db8616a-4185-4b5f-84d2-41b876b0ac27",
          "game_id": "sp80",
          "game_version": "589a99af",
          "published_at": "2026-09-16T04:34:42.01718Z",
          "score": 100,
          "model": "gpt-6-1-sol",
          "runner": "benchmark_agent",
          "config": "openai-gpt-6-1-sol-high-provider-adapter",
          "verified_publish": true,
          "levels_completed": 6,
          "actions": 187,
          "resets": 0,
          "state": "WIN",
          "completed": true,
          "number_of_levels": 0,
          "level_scores": [
            115,
            115,
            115,
            115,
            115,
            115
          ],
          "level_actions": [
            10,
            9,
            22,
            65,
            38,
            43
          ],
          "level_baseline_actions": [
            39,
            58,
            25,
            148,
            96,
            152
          ]
        },
        {
          "session_id": "db32913f-7998-470c-a6de-a98b1f8ab630",
          "game_id": "sp80",
          "game_version": "589a99af",
          "published_at": "2026-09-16T03:42:04.798061Z",
          "score": 98.81438289601556,
          "model": "gpt-6-1-sol",
          "runner": "benchmark_agent",
          "config": "openai-gpt-6-1-sol-xhigh-provider-adapter",
          "verified_publish": true,
          "levels_completed": 6,
          "actions": 259,
          "resets": 1,
          "state": "WIN",
          "completed": true,
          "number_of_levels": 0,
          "level_scores": [
            115,
            115,
            115,
            115,
            47.02040816326531,
            115
          ],
          "level_actions": [
            9,
            7,
            11,
            45,
            140,
            47
          ],
          "level_baseline_actions": [
            39,
            58,
            25,
            148,
            96,
            152
          ]
        },
        {
          "session_id": "e64ddcf5-8a45-4321-b851-b012dac0daac",
          "game_id": "sp80",
          "game_version": "589a99af",
          "published_at": "2026-09-16T01:23:17.015092Z",
          "score": 100,
          "model": "gpt-6-1-sol",
          "runner": "benchmark_agent",
          "config": "openai-gpt-6-1-sol-max-provider-adapter",
          "verified_publish": true,
          "levels_completed": 6,
          "actions": 188,
          "resets": 1,
          "state": "WIN",
          "completed": true,
          "number_of_levels": 0,
          "level_scores": [
            115,
            115,
            115,
            115,
            115,
            115
          ],
          "level_actions": [
            9,
            9,
            17,
            68,
            46,
            39
          ],
          "level_baseline_actions": [
            39,
            58,
            25,
            148,
            96,
            152
          ]
        },
        {
          "session_id": "39843f76-2ab5-49be-8b13-f5929125615b",
          "game_id": "sp80",
          "game_version": "589a99af",
          "published_at": "2026-09-15T21:28:21.398391Z",
          "score": 100,
          "model": "gpt-6-1-sol",
          "runner": "benchmark_agent",
          "config": "openai-gpt-6-1-sol-max",
          "verified_publish": true,
          "levels_completed": 6,
          "actions": 162,
          "resets": 0,
          "state": "WIN",
          "completed": true,
          "number_of_levels": 0,
          "level_scores": [
            115,
            115,
            115,
            115,
            115,
            115
          ],
          "level_actions": [
            5,
            7,
            11,
            63,
            36,
            40
          ],
          "level_baseline_actions": [
            39,
            58,
            25,
            148,
            96,
            152
          ]
        },
        {
          "session_id": "f0f73094-c37f-4e2f-910c-7ad1a2757b77",
          "game_id": "sp80",
          "game_version": "589a99af",
          "published_at": "2026-09-11T20:02:05.318215Z",
          "score": 14.285714285714285,
          "model": "gemini-3.8-flash",
          "runner": "benchmark_agent",
          "config": "google-gemini-3-8-flash-medium",
          "verified_publish": true,
          "levels_completed": 2,
          "actions": 159,
          "resets": 8,
          "state": "GAME_OVER",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            115,
            115,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            17,
            17,
            125,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            39,
            58,
            25,
            148,
            96,
            152
          ]
        },
        {
          "session_id": "75876e47-e8d3-40b9-82f8-8695e07cb711",
          "game_id": "sp80",
          "game_version": "589a99af",
          "published_at": "2026-09-09T17:33:04.546121Z",
          "score": 14.285714285714285,
          "model": "gemini-3.8-flash",
          "runner": "benchmark_agent",
          "config": "google-gemini-3-8-flash-low",
          "verified_publish": true,
          "levels_completed": 2,
          "actions": 36,
          "resets": 0,
          "state": "NOT_FINISHED",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            115,
            115,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            7,
            17,
            12,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            39,
            58,
            25,
            148,
            96,
            152
          ]
        },
        {
          "session_id": "b52edbfb-532a-4117-a505-988006b8c6ea",
          "game_id": "sp80",
          "game_version": "589a99af",
          "published_at": "2026-09-09T17:06:45.418378Z",
          "score": 14.285714285714285,
          "model": "gemini-3.8-flash",
          "runner": "benchmark_agent",
          "config": "google-gemini-3-8-flash-high",
          "verified_publish": true,
          "levels_completed": 2,
          "actions": 150,
          "resets": 2,
          "state": "GAME_OVER",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            115,
            115,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            13,
            12,
            125,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            39,
            58,
            25,
            148,
            96,
            152
          ]
        },
        {
          "session_id": "5ed533eb-6f62-4914-ad9a-354476491e94",
          "game_id": "sp80",
          "game_version": "589a99af",
          "published_at": "2026-08-11T09:05:01.804257Z",
          "score": 4.761904761904762,
          "model": "grok-4.6",
          "runner": "benchmark_agent",
          "config": "xai-grok-4-6-xhigh",
          "verified_publish": true,
          "levels_completed": 1,
          "actions": 328,
          "resets": 9,
          "state": "GAME_OVER",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            105.33240997229917,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            38,
            290,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            39,
            58,
            25,
            148,
            96,
            152
          ]
        },
        {
          "session_id": "3f9fcf5d-33e7-4bb2-bfc5-055433c573b2",
          "game_id": "sp80",
          "game_version": "589a99af",
          "published_at": "2026-07-22T13:02:59.085687Z",
          "score": 56.305773862711575,
          "model": "claude-opus-5",
          "runner": "benchmark_agent",
          "config": "anthropic-claude-opus-5-high",
          "verified_publish": true,
          "levels_completed": 5,
          "actions": 669,
          "resets": 9,
          "state": "GAME_OVER",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            115,
            115,
            115,
            115,
            6.484250223388612,
            0
          ],
          "level_actions": [
            10,
            9,
            12,
            131,
            377,
            130
          ],
          "level_baseline_actions": [
            39,
            58,
            25,
            148,
            96,
            152
          ]
        },
        {
          "session_id": "b32d9429-3dc0-4103-bb51-f2d9747a9a48",
          "game_id": "sp80",
          "game_version": "589a99af",
          "published_at": "2026-06-24T10:51:14.31261Z",
          "score": 0,
          "model": "gpt-5-6-luna",
          "runner": "benchmark_agent",
          "config": "openai-gpt-5-6-luna-low",
          "verified_publish": true,
          "levels_completed": 0,
          "actions": 195,
          "resets": 6,
          "state": "GAME_OVER",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            195,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            39,
            58,
            25,
            148,
            96,
            152
          ]
        },
        {
          "session_id": "6f4c3dc5-73c6-44bc-abcf-308f11cfd618",
          "game_id": "sp80",
          "game_version": "589a99af",
          "published_at": "2026-06-24T10:29:43.0085Z",
          "score": 0,
          "model": "gpt-5-6-luna",
          "runner": "benchmark_agent",
          "config": "openai-gpt-5-6-luna-medium",
          "verified_publish": true,
          "levels_completed": 0,
          "actions": 195,
          "resets": 7,
          "state": "GAME_OVER",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            195,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            39,
            58,
            25,
            148,
            96,
            152
          ]
        },
        {
          "session_id": "e9db3405-5d6c-405a-a65f-15c503b62227",
          "game_id": "sp80",
          "game_version": "589a99af",
          "published_at": "2026-06-24T03:54:42.618225Z",
          "score": 0,
          "model": "gpt-5-6-luna",
          "runner": "benchmark_agent",
          "config": "openai-gpt-5-6-luna-high",
          "verified_publish": true,
          "levels_completed": 0,
          "actions": 195,
          "resets": 16,
          "state": "GAME_OVER",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            195,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            39,
            58,
            25,
            148,
            96,
            152
          ]
        },
        {
          "session_id": "bd09be28-2163-4b2f-a708-b72a329aca60",
          "game_id": "sp80",
          "game_version": "589a99af",
          "published_at": "2026-06-24T00:57:40.362819Z",
          "score": 0,
          "model": "gpt-5-6-luna",
          "runner": "benchmark_agent",
          "config": "openai-gpt-5-6-luna-xhigh",
          "verified_publish": true,
          "levels_completed": 0,
          "actions": 195,
          "resets": 25,
          "state": "NOT_FINISHED",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            195,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            39,
            58,
            25,
            148,
            96,
            152
          ]
        },
        {
          "session_id": "ba888030-2e19-4b28-80fc-c5332f470f23",
          "game_id": "sp80",
          "game_version": "589a99af",
          "published_at": "2026-06-23T01:07:36.36917Z",
          "score": 0,
          "model": "gpt-5-6-luna",
          "runner": "benchmark_agent",
          "config": "openai-gpt-5-6-luna-max",
          "verified_publish": true,
          "levels_completed": 0,
          "actions": 195,
          "resets": 53,
          "state": "NOT_FINISHED",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            195,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            39,
            58,
            25,
            148,
            96,
            152
          ]
        },
        {
          "session_id": "b8b2189d-274f-4dc4-aa19-cda99f14a5e3",
          "game_id": "sp80",
          "game_version": "589a99af",
          "published_at": "2026-06-20T12:02:27.621413Z",
          "score": 0,
          "model": "gpt-5-6-terra",
          "runner": "benchmark_agent",
          "config": "openai-gpt-5-6-terra-low",
          "verified_publish": true,
          "levels_completed": 0,
          "actions": 195,
          "resets": 6,
          "state": "GAME_OVER",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            195,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            39,
            58,
            25,
            148,
            96,
            152
          ]
        },
        {
          "session_id": "2a871ed7-67b7-45fe-8ee1-ad46a0208722",
          "game_id": "sp80",
          "game_version": "589a99af",
          "published_at": "2026-06-20T09:40:32.527595Z",
          "score": 0.26929123820854933,
          "model": "gpt-5-6-terra",
          "runner": "benchmark_agent",
          "config": "openai-gpt-5-6-terra-medium",
          "verified_publish": true,
          "levels_completed": 1,
          "actions": 454,
          "resets": 14,
          "state": "GAME_OVER",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            5.655116002379536,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            164,
            290,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            39,
            58,
            25,
            148,
            96,
            152
          ]
        },
        {
          "session_id": "bef72cae-b35b-4f32-84e3-63904f9fa245",
          "game_id": "sp80",
          "game_version": "589a99af",
          "published_at": "2026-06-20T09:17:03.396258Z",
          "score": 1.5212890449185341,
          "model": "gpt-5-6-terra",
          "runner": "benchmark_agent",
          "config": "openai-gpt-5-6-terra-high",
          "verified_publish": true,
          "levels_completed": 1,
          "actions": 359,
          "resets": 11,
          "state": "GAME_OVER",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            31.94706994328922,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            69,
            290,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            39,
            58,
            25,
            148,
            96,
            152
          ]
        },
        {
          "session_id": "7644dcb9-7f96-4ce2-a331-abed22e97e03",
          "game_id": "sp80",
          "game_version": "589a99af",
          "published_at": "2026-06-20T04:09:32.410952Z",
          "score": 4.761904761904762,
          "model": "gpt-5-6-terra",
          "runner": "benchmark_agent",
          "config": "openai-gpt-5-6-terra-xhigh",
          "verified_publish": true,
          "levels_completed": 1,
          "actions": 323,
          "resets": 11,
          "state": "GAME_OVER",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            115,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            33,
            290,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            39,
            58,
            25,
            148,
            96,
            152
          ]
        },
        {
          "session_id": "c34935e7-1ef8-4296-8dec-66c61b4c8ca5",
          "game_id": "sp80",
          "game_version": "589a99af",
          "published_at": "2026-06-19T00:14:41.782745Z",
          "score": 4.761904761904762,
          "model": "gpt-5-6-terra",
          "runner": "benchmark_agent",
          "config": "openai-gpt-5-6-terra-max",
          "verified_publish": true,
          "levels_completed": 1,
          "actions": 302,
          "resets": 12,
          "state": "GAME_OVER",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            115,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            12,
            290,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            39,
            58,
            25,
            148,
            96,
            152
          ]
        },
        {
          "session_id": "6f9b96ba-d1bc-4fc2-b42f-99e8f1018b78",
          "game_id": "sp80",
          "game_version": "589a99af",
          "published_at": "2026-06-13T19:53:18.913185Z",
          "score": 28.57142857142857,
          "model": "gpt-5-6-sol",
          "runner": "benchmark_agent",
          "config": "openai-gpt-5-6-sol-max",
          "verified_publish": true,
          "levels_completed": 3,
          "actions": 779,
          "resets": 21,
          "state": "GAME_OVER",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            115,
            115,
            115,
            0,
            0,
            0
          ],
          "level_actions": [
            5,
            11,
            23,
            740,
            0,
            0
          ],
          "level_baseline_actions": [
            39,
            58,
            25,
            148,
            96,
            152
          ]
        },
        {
          "session_id": "acb1e3d0-ea22-4e07-8c96-d8a06ce1ed26",
          "game_id": "sp80",
          "game_version": "589a99af",
          "published_at": "2026-06-12T00:31:13.762742Z",
          "score": 28.57142857142857,
          "model": "gpt-5-6-sol",
          "runner": "benchmark_agent",
          "config": "openai-gpt-5-6-sol-xhigh",
          "verified_publish": true,
          "levels_completed": 3,
          "actions": 554,
          "resets": 14,
          "state": "GAME_OVER",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            115,
            115,
            115,
            0,
            0,
            0
          ],
          "level_actions": [
            7,
            30,
            10,
            507,
            0,
            0
          ],
          "level_baseline_actions": [
            39,
            58,
            25,
            148,
            96,
            152
          ]
        },
        {
          "session_id": "f4556b8c-6b93-4202-baf2-ba50949dcc0d",
          "game_id": "sp80",
          "game_version": "589a99af",
          "published_at": "2026-06-10T11:36:14.398149Z",
          "score": 14.285714285714285,
          "model": "gpt-5-6-sol",
          "runner": "benchmark_agent",
          "config": "openai-gpt-5-6-sol-high",
          "verified_publish": true,
          "levels_completed": 2,
          "actions": 145,
          "resets": 5,
          "state": "GAME_OVER",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            115,
            115,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            9,
            11,
            125,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            39,
            58,
            25,
            148,
            96,
            152
          ]
        },
        {
          "session_id": "cc95ab20-eb64-476f-bebf-718a614b6967",
          "game_id": "sp80",
          "game_version": "589a99af",
          "published_at": "2026-06-10T08:51:06.970252Z",
          "score": 4.761904761904762,
          "model": "gpt-5-6-sol",
          "runner": "benchmark_agent",
          "config": "openai-gpt-5-6-sol-medium",
          "verified_publish": true,
          "levels_completed": 1,
          "actions": 324,
          "resets": 15,
          "state": "GAME_OVER",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            115,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            34,
            290,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            39,
            58,
            25,
            148,
            96,
            152
          ]
        },
        {
          "session_id": "ed6bcdc3-cf5c-4798-b07c-3974f494236a",
          "game_id": "sp80",
          "game_version": "589a99af",
          "published_at": "2026-06-09T22:05:24.370629Z",
          "score": 4.761904761904762,
          "model": "gpt-5-6-sol",
          "runner": "benchmark_agent",
          "config": "openai-gpt-5-6-sol-low",
          "verified_publish": true,
          "levels_completed": 1,
          "actions": 301,
          "resets": 11,
          "state": "GAME_OVER",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            115,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            11,
            290,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            39,
            58,
            25,
            148,
            96,
            152
          ]
        },
        {
          "session_id": "556fef38-a1e3-4d4e-82c2-0148443f974e",
          "game_id": "sp80",
          "game_version": "589a99af",
          "published_at": "2026-05-29T20:09:56.034006Z",
          "score": 14.285714285714285,
          "model": "claude-opus-4-8",
          "runner": "benchmark_agent",
          "config": "anthropic-opus-4-8-high",
          "verified_publish": true,
          "levels_completed": 2,
          "actions": 178,
          "resets": 11,
          "state": "GAME_OVER",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            115,
            115,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            35,
            18,
            125,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            39,
            58,
            25,
            148,
            96,
            152
          ]
        },
        {
          "session_id": "d1a3f504-3924-481f-bf97-1b4d09528de9",
          "game_id": "sp80",
          "game_version": "589a99af",
          "published_at": "2026-05-01T05:08:04.405095Z",
          "score": 0,
          "model": "claude-opus-4-7",
          "runner": "benchmark_agent",
          "config": "anthropic-opus-4-7-high",
          "verified_publish": true,
          "levels_completed": 0,
          "actions": 195,
          "resets": 7,
          "state": "GAME_OVER",
          "completed": false,
          "number_of_levels": 6,
          "level_scores": [
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            195,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            39,
            58,
            25,
            148,
            96,
            152
          ]
        },
        {
          "session_id": "e2f333be-8318-449e-9217-3be9413b9933",
          "game_id": "sp80",
          "game_version": "589a99af",
          "published_at": "2026-04-30T23:22:19.160554Z",
          "score": 4.761904761904762,
          "model": "gpt-5.5-2026-04-23",
          "runner": "benchmark_agent",
          "config": "openai-gpt-5-5-2026-04-23-high",
          "verified_publish": true,
          "levels_completed": 1,
          "actions": 298,
          "resets": 9,
          "state": "GAME_OVER",
          "completed": false,
          "number_of_levels": 6,
          "level_scores": [
            115,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            8,
            290,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            39,
            58,
            25,
            148,
            96,
            152
          ]
        },
        {
          "session_id": "a8a3bfa2-33e0-41e0-ab28-04b7dc87c171",
          "game_id": "sp80",
          "game_version": "589a99af",
          "published_at": "2026-04-14T17:44:17.517867Z",
          "score": 0,
          "model": "grok-4.20-beta-0309-reasoning",
          "runner": "benchmark_agent",
          "config": "xai-grok-4-20-beta-0309-reasoning",
          "verified_publish": true,
          "levels_completed": 0,
          "actions": 133,
          "resets": 5,
          "state": "GAME_OVER",
          "completed": false,
          "number_of_levels": 6,
          "level_scores": [
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            133,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            39,
            58,
            25,
            148,
            96,
            152
          ]
        },
        {
          "session_id": "8128b3bc-01a7-442d-893e-44f73326e117",
          "game_id": "sp80",
          "game_version": "589a99af",
          "published_at": "2026-04-14T17:44:16.825051Z",
          "score": 1.3591400155483475,
          "model": "gpt-5.4-2026-03-05",
          "runner": "benchmark_agent",
          "config": "openai-gpt-5-4-2026-03-05-high",
          "verified_publish": true,
          "levels_completed": 1,
          "actions": 363,
          "resets": 8,
          "state": "GAME_OVER",
          "completed": false,
          "number_of_levels": 6,
          "level_scores": [
            28.5419403265153,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            73,
            290,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            39,
            58,
            25,
            148,
            96,
            152
          ]
        },
        {
          "session_id": "c3a90fd3-bc10-49ab-b2aa-bb997ce57121",
          "game_id": "sp80",
          "game_version": "589a99af",
          "published_at": "2026-04-14T17:44:15.509183Z",
          "score": 4.761904761904762,
          "model": "gemini-3.1-pro-preview",
          "runner": "benchmark_agent",
          "config": "google-gemini-3-1-pro-preview",
          "verified_publish": true,
          "levels_completed": 1,
          "actions": 51,
          "resets": 1,
          "state": "GAME_OVER",
          "completed": false,
          "number_of_levels": 6,
          "level_scores": [
            115,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            5,
            46,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            39,
            58,
            25,
            148,
            96,
            152
          ]
        },
        {
          "session_id": "1c4e5367-29f6-4c67-b63f-753280e15edf",
          "game_id": "sp80",
          "game_version": "589a99af",
          "published_at": "2026-04-14T17:44:14.807649Z",
          "score": 4.761904761904762,
          "model": "claude-opus-4-6",
          "runner": "benchmark_agent",
          "config": "anthropic-opus-4-6-max-effort",
          "verified_publish": true,
          "levels_completed": 1,
          "actions": 90,
          "resets": 2,
          "state": "GAME_OVER",
          "completed": false,
          "number_of_levels": 6,
          "level_scores": [
            115,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            18,
            72,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            39,
            58,
            25,
            148,
            96,
            152
          ]
        }
      ]
    },
    {
      "game": "tr87",
      "source": "https://arcprize.org/tasks/tr87",
      "runs": [
        {
          "session_id": "a76326cf-1087-47ec-bd53-74e667a25859",
          "game_id": "tr87",
          "game_version": "cd924810",
          "published_at": "2026-09-26T03:33:25.626163Z",
          "score": 0,
          "model": "gpt-6-sol",
          "runner": "benchmark_agent",
          "config": "openai-gpt-6-sol-max",
          "verified_publish": true,
          "levels_completed": 0,
          "actions": 157,
          "resets": 1,
          "state": "NOT_FINISHED",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            157,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            54,
            58,
            40,
            45,
            71,
            146
          ]
        },
        {
          "session_id": "dd81c927-3113-4cba-8679-2ae6cc9a85a1",
          "game_id": "tr87",
          "game_version": "cd924810",
          "published_at": "2026-09-25T13:10:56.914964Z",
          "score": 0,
          "model": "gpt-6-sol",
          "runner": "benchmark_agent",
          "config": "openai-gpt-6-sol-none",
          "verified_publish": true,
          "levels_completed": 0,
          "actions": 270,
          "resets": 2,
          "state": "GAME_OVER",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            270,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            54,
            58,
            40,
            45,
            71,
            146
          ]
        },
        {
          "session_id": "fbc212f9-b39e-4138-a026-824674ac7d3d",
          "game_id": "tr87",
          "game_version": "cd924810",
          "published_at": "2026-09-25T12:12:00.812178Z",
          "score": 0,
          "model": "gpt-6-sol",
          "runner": "benchmark_agent",
          "config": "openai-gpt-6-sol-xhigh",
          "verified_publish": true,
          "levels_completed": 0,
          "actions": 270,
          "resets": 4,
          "state": "NOT_FINISHED",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            270,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            54,
            58,
            40,
            45,
            71,
            146
          ]
        },
        {
          "session_id": "b3679c95-4f16-4497-b7ad-91ad816951da",
          "game_id": "tr87",
          "game_version": "cd924810",
          "published_at": "2026-09-25T10:33:22.234714Z",
          "score": 30.147854280643266,
          "model": "gpt-6-sol",
          "runner": "benchmark_agent",
          "config": "openai-gpt-6-sol-xhigh-provider-adapter",
          "verified_publish": true,
          "levels_completed": 5,
          "actions": 1166,
          "resets": 3,
          "state": "GAME_OVER",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            115,
            115,
            22.675736961451246,
            40.17060107121603,
            11.87906494485814,
            0
          ],
          "level_actions": [
            43,
            32,
            84,
            71,
            206,
            730
          ],
          "level_baseline_actions": [
            54,
            58,
            40,
            45,
            71,
            146
          ]
        },
        {
          "session_id": "d423b619-ac10-4a21-814f-6515083d5769",
          "game_id": "tr87",
          "game_version": "cd924810",
          "published_at": "2026-09-25T09:23:41.125901Z",
          "score": 0,
          "model": "gpt-6-sol",
          "runner": "benchmark_agent",
          "config": "openai-gpt-6-sol-high",
          "verified_publish": true,
          "levels_completed": 0,
          "actions": 270,
          "resets": 5,
          "state": "NOT_FINISHED",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            270,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            54,
            58,
            40,
            45,
            71,
            146
          ]
        },
        {
          "session_id": "4ed9260b-bc52-4db7-88c1-ca712b7af755",
          "game_id": "tr87",
          "game_version": "cd924810",
          "published_at": "2026-09-25T07:16:36.010851Z",
          "score": 0.503908923127968,
          "model": "gpt-6-sol",
          "runner": "benchmark_agent",
          "config": "openai-gpt-6-sol-medium-provider-adapter",
          "verified_publish": true,
          "levels_completed": 1,
          "actions": 456,
          "resets": 4,
          "state": "NOT_FINISHED",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            10.582087385687327,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            166,
            290,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            54,
            58,
            40,
            45,
            71,
            146
          ]
        },
        {
          "session_id": "c534f5c1-0b78-48c7-9d06-7872b8329d3e",
          "game_id": "tr87",
          "game_version": "cd924810",
          "published_at": "2026-09-25T06:55:41.469166Z",
          "score": 0,
          "model": "gpt-6-sol",
          "runner": "benchmark_agent",
          "config": "openai-gpt-6-sol-medium",
          "verified_publish": true,
          "levels_completed": 0,
          "actions": 270,
          "resets": 3,
          "state": "GAME_OVER",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            270,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            54,
            58,
            40,
            45,
            71,
            146
          ]
        },
        {
          "session_id": "d9339a22-3f3e-48ab-8170-027339ce8362",
          "game_id": "tr87",
          "game_version": "cd924810",
          "published_at": "2026-09-25T06:52:20.069711Z",
          "score": 0,
          "model": "gpt-6-sol",
          "runner": "benchmark_agent",
          "config": "openai-gpt-6-sol-none-provider-adapter",
          "verified_publish": true,
          "levels_completed": 0,
          "actions": 270,
          "resets": 3,
          "state": "NOT_FINISHED",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            270,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            54,
            58,
            40,
            45,
            71,
            146
          ]
        },
        {
          "session_id": "a06cdce7-400d-425f-88d1-16079547e470",
          "game_id": "tr87",
          "game_version": "cd924810",
          "published_at": "2026-09-25T06:40:20.706002Z",
          "score": 0,
          "model": "gpt-6-sol",
          "runner": "benchmark_agent",
          "config": "openai-gpt-6-sol-low-provider-adapter",
          "verified_publish": true,
          "levels_completed": 0,
          "actions": 270,
          "resets": 3,
          "state": "NOT_FINISHED",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            270,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            54,
            58,
            40,
            45,
            71,
            146
          ]
        },
        {
          "session_id": "43846c18-6334-4542-aa57-cd371a1fd746",
          "game_id": "tr87",
          "game_version": "cd924810",
          "published_at": "2026-09-25T06:21:16.00627Z",
          "score": 0,
          "model": "gpt-6-sol",
          "runner": "benchmark_agent",
          "config": "openai-gpt-6-sol-low",
          "verified_publish": true,
          "levels_completed": 0,
          "actions": 270,
          "resets": 2,
          "state": "GAME_OVER",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            270,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            54,
            58,
            40,
            45,
            71,
            146
          ]
        },
        {
          "session_id": "a42fd6c1-f7ab-4c15-a6a8-689f2d3ac9e7",
          "game_id": "tr87",
          "game_version": "cd924810",
          "published_at": "2026-09-25T05:47:59.267324Z",
          "score": 14.285714285714285,
          "model": "gpt-6-sol",
          "runner": "benchmark_agent",
          "config": "openai-gpt-6-sol-high-provider-adapter",
          "verified_publish": true,
          "levels_completed": 2,
          "actions": 281,
          "resets": 1,
          "state": "NOT_FINISHED",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            115,
            115,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            40,
            41,
            200,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            54,
            58,
            40,
            45,
            71,
            146
          ]
        },
        {
          "session_id": "7213cf48-af16-45f5-8b9c-6b69abba97e4",
          "game_id": "tr87",
          "game_version": "cd924810",
          "published_at": "2026-09-21T18:07:02.547672Z",
          "score": 4.127739086121964,
          "model": "gemini-3.8-flash",
          "runner": "benchmark_agent",
          "config": "google-gemini-3-8-flash-low-provider-adapter",
          "verified_publish": true,
          "levels_completed": 1,
          "actions": 348,
          "resets": 3,
          "state": "GAME_OVER",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            86.68252080856124,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            58,
            290,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            54,
            58,
            40,
            45,
            71,
            146
          ]
        },
        {
          "session_id": "3c9292e4-c882-42d1-9de2-e4c80bc752f7",
          "game_id": "tr87",
          "game_version": "cd924810",
          "published_at": "2026-09-20T11:35:53.55291Z",
          "score": 4.761904761904762,
          "model": "gemini-3.8-flash",
          "runner": "benchmark_agent",
          "config": "google-gemini-3-8-flash-medium-provider-adapter",
          "verified_publish": true,
          "levels_completed": 1,
          "actions": 331,
          "resets": 2,
          "state": "GAME_OVER",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            115,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            41,
            290,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            54,
            58,
            40,
            45,
            71,
            146
          ]
        },
        {
          "session_id": "d966256d-f27f-42dd-8c51-03b74ffb2b5c",
          "game_id": "tr87",
          "game_version": "cd924810",
          "published_at": "2026-09-18T00:24:12.896596Z",
          "score": 4.761904761904762,
          "model": "gemini-3.8-flash",
          "runner": "benchmark_agent",
          "config": "google-gemini-3-8-flash-high-provider-adapter",
          "verified_publish": true,
          "levels_completed": 1,
          "actions": 321,
          "resets": 2,
          "state": "GAME_OVER",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            115,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            31,
            290,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            54,
            58,
            40,
            45,
            71,
            146
          ]
        },
        {
          "session_id": "24460385-fa2b-4371-bb81-33432802afd1",
          "game_id": "tr87",
          "game_version": "cd924810",
          "published_at": "2026-09-16T19:42:30.421942Z",
          "score": 0,
          "model": "gpt-6-luna",
          "runner": "benchmark_agent",
          "config": "openai-gpt-6-luna-medium-provider-adapter",
          "verified_publish": true,
          "levels_completed": 0,
          "actions": 270,
          "resets": 2,
          "state": "GAME_OVER",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            270,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            54,
            58,
            40,
            45,
            71,
            146
          ]
        },
        {
          "session_id": "98d10102-d53a-4610-b5dc-d97f22fb2e23",
          "game_id": "tr87",
          "game_version": "cd924810",
          "published_at": "2026-09-16T19:35:16.141098Z",
          "score": 0,
          "model": "gpt-6-luna",
          "runner": "benchmark_agent",
          "config": "openai-gpt-6-luna-low",
          "verified_publish": true,
          "levels_completed": 0,
          "actions": 270,
          "resets": 2,
          "state": "GAME_OVER",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            270,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            54,
            58,
            40,
            45,
            71,
            146
          ]
        },
        {
          "session_id": "e935c1e4-65e8-428c-bd63-5940e3716011",
          "game_id": "tr87",
          "game_version": "cd924810",
          "published_at": "2026-09-16T19:30:03.929377Z",
          "score": 0,
          "model": "gpt-6-luna",
          "runner": "benchmark_agent",
          "config": "openai-gpt-6-luna-low-provider-adapter",
          "verified_publish": true,
          "levels_completed": 0,
          "actions": 270,
          "resets": 3,
          "state": "GAME_OVER",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            270,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            54,
            58,
            40,
            45,
            71,
            146
          ]
        },
        {
          "session_id": "549704c5-1f8c-4f8a-b9c4-8b4752041d2a",
          "game_id": "tr87",
          "game_version": "cd924810",
          "published_at": "2026-09-16T19:27:44.823308Z",
          "score": 0,
          "model": "gpt-6-luna",
          "runner": "benchmark_agent",
          "config": "openai-gpt-6-luna-none",
          "verified_publish": true,
          "levels_completed": 0,
          "actions": 270,
          "resets": 7,
          "state": "NOT_FINISHED",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            270,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            54,
            58,
            40,
            45,
            71,
            146
          ]
        },
        {
          "session_id": "c4cb84bf-e138-4116-95d3-6c486424d433",
          "game_id": "tr87",
          "game_version": "cd924810",
          "published_at": "2026-09-16T16:41:09.722598Z",
          "score": 0,
          "model": "gpt-6-luna",
          "runner": "benchmark_agent",
          "config": "openai-gpt-6-luna-high",
          "verified_publish": true,
          "levels_completed": 0,
          "actions": 270,
          "resets": 3,
          "state": "NOT_FINISHED",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            270,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            54,
            58,
            40,
            45,
            71,
            146
          ]
        },
        {
          "session_id": "a760f0d3-21e0-441d-9517-139a8c60b741",
          "game_id": "tr87",
          "game_version": "cd924810",
          "published_at": "2026-09-16T14:56:17.023854Z",
          "score": 0,
          "model": "gpt-6-luna",
          "runner": "benchmark_agent",
          "config": "openai-gpt-6-luna-none-provider-adapter",
          "verified_publish": true,
          "levels_completed": 0,
          "actions": 270,
          "resets": 5,
          "state": "GAME_OVER",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            270,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            54,
            58,
            40,
            45,
            71,
            146
          ]
        },
        {
          "session_id": "a3a05453-faf9-4507-b105-221c718b84ec",
          "game_id": "tr87",
          "game_version": "cd924810",
          "published_at": "2026-09-16T13:06:09.565682Z",
          "score": 0,
          "model": "gpt-6-luna",
          "runner": "benchmark_agent",
          "config": "openai-gpt-6-luna-max",
          "verified_publish": true,
          "levels_completed": 0,
          "actions": 270,
          "resets": 3,
          "state": "NOT_FINISHED",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            270,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            54,
            58,
            40,
            45,
            71,
            146
          ]
        },
        {
          "session_id": "624a1479-3178-44ea-b57a-3813d285ab40",
          "game_id": "tr87",
          "game_version": "cd924810",
          "published_at": "2026-09-16T09:24:30.167137Z",
          "score": 0,
          "model": "gpt-6-luna",
          "runner": "benchmark_agent",
          "config": "openai-gpt-6-luna-medium",
          "verified_publish": true,
          "levels_completed": 0,
          "actions": 270,
          "resets": 2,
          "state": "NOT_FINISHED",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            270,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            54,
            58,
            40,
            45,
            71,
            146
          ]
        },
        {
          "session_id": "64e92714-1348-450d-9576-69b7474ce63b",
          "game_id": "tr87",
          "game_version": "cd924810",
          "published_at": "2026-09-16T09:05:42.180479Z",
          "score": 0,
          "model": "gpt-6-luna",
          "runner": "benchmark_agent",
          "config": "openai-gpt-6-luna-high-provider-adapter",
          "verified_publish": true,
          "levels_completed": 0,
          "actions": 270,
          "resets": 2,
          "state": "GAME_OVER",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            270,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            54,
            58,
            40,
            45,
            71,
            146
          ]
        },
        {
          "session_id": "5a9a0911-07db-4b06-8999-9fc12938e3a4",
          "game_id": "tr87",
          "game_version": "cd924810",
          "published_at": "2026-09-16T08:34:27.972427Z",
          "score": 0,
          "model": "gpt-6-luna",
          "runner": "benchmark_agent",
          "config": "openai-gpt-6-luna-xhigh",
          "verified_publish": true,
          "levels_completed": 0,
          "actions": 270,
          "resets": 2,
          "state": "GAME_OVER",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            270,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            54,
            58,
            40,
            45,
            71,
            146
          ]
        },
        {
          "session_id": "5c13ffc5-9094-4c6e-9b57-ec6a59cde3bc",
          "game_id": "tr87",
          "game_version": "cd924810",
          "published_at": "2026-09-16T08:30:51.294016Z",
          "score": 0,
          "model": "gpt-6-luna",
          "runner": "benchmark_agent",
          "config": "openai-gpt-6-luna-xhigh-provider-adapter",
          "verified_publish": true,
          "levels_completed": 0,
          "actions": 270,
          "resets": 2,
          "state": "GAME_OVER",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            270,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            54,
            58,
            40,
            45,
            71,
            146
          ]
        },
        {
          "session_id": "26cf7c9c-b14c-4206-be3e-cf3af1179557",
          "game_id": "tr87",
          "game_version": "cd924810",
          "published_at": "2026-09-16T07:20:10.400573Z",
          "score": 0,
          "model": "gpt-6-1-sol",
          "runner": "benchmark_agent",
          "config": "openai-gpt-6-1-sol-low",
          "verified_publish": true,
          "levels_completed": 0,
          "actions": 270,
          "resets": 2,
          "state": "GAME_OVER",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            270,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            54,
            58,
            40,
            45,
            71,
            146
          ]
        },
        {
          "session_id": "1e12982e-2bef-453f-bb38-1355b38beffc",
          "game_id": "tr87",
          "game_version": "cd924810",
          "published_at": "2026-09-16T06:18:05.838361Z",
          "score": 4.761904761904762,
          "model": "gpt-6-1-sol",
          "runner": "benchmark_agent",
          "config": "openai-gpt-6-1-sol-medium",
          "verified_publish": true,
          "levels_completed": 1,
          "actions": 328,
          "resets": 2,
          "state": "GAME_OVER",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            115,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            38,
            290,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            54,
            58,
            40,
            45,
            71,
            146
          ]
        },
        {
          "session_id": "f3261b72-2840-498f-8cac-e284640112bd",
          "game_id": "tr87",
          "game_version": "cd924810",
          "published_at": "2026-09-16T05:31:53.51515Z",
          "score": 100,
          "model": "gpt-6-1-sol",
          "runner": "benchmark_agent",
          "config": "openai-gpt-6-1-sol-xhigh",
          "verified_publish": true,
          "levels_completed": 6,
          "actions": 220,
          "resets": 0,
          "state": "WIN",
          "completed": true,
          "number_of_levels": 0,
          "level_scores": [
            115,
            115,
            115,
            115,
            115,
            115
          ],
          "level_actions": [
            41,
            28,
            26,
            28,
            43,
            54
          ],
          "level_baseline_actions": [
            54,
            58,
            40,
            45,
            71,
            146
          ]
        },
        {
          "session_id": "aa1de804-a14b-4e2c-8897-2bd92fcb5f50",
          "game_id": "tr87",
          "game_version": "cd924810",
          "published_at": "2026-09-16T05:25:07.930551Z",
          "score": 100,
          "model": "gpt-6-1-sol",
          "runner": "benchmark_agent",
          "config": "openai-gpt-6-1-sol-high",
          "verified_publish": true,
          "levels_completed": 6,
          "actions": 338,
          "resets": 1,
          "state": "WIN",
          "completed": true,
          "number_of_levels": 0,
          "level_scores": [
            89.75069252077562,
            115,
            110.80332409972297,
            115,
            115,
            104.23981612792804
          ],
          "level_actions": [
            57,
            29,
            38,
            29,
            42,
            143
          ],
          "level_baseline_actions": [
            54,
            58,
            40,
            45,
            71,
            146
          ]
        },
        {
          "session_id": "4627782b-d8c3-462b-bb92-93cf6ea5d099",
          "game_id": "tr87",
          "game_version": "cd924810",
          "published_at": "2026-09-16T04:51:57.692276Z",
          "score": 100,
          "model": "gpt-6-1-sol",
          "runner": "benchmark_agent",
          "config": "openai-gpt-6-1-sol-high-provider-adapter",
          "verified_publish": true,
          "levels_completed": 6,
          "actions": 165,
          "resets": 0,
          "state": "WIN",
          "completed": true,
          "number_of_levels": 0,
          "level_scores": [
            115,
            115,
            115,
            115,
            115,
            115
          ],
          "level_actions": [
            23,
            32,
            27,
            21,
            29,
            33
          ],
          "level_baseline_actions": [
            54,
            58,
            40,
            45,
            71,
            146
          ]
        },
        {
          "session_id": "0c897998-bf91-46ac-8ff7-f3921d656964",
          "game_id": "tr87",
          "game_version": "cd924810",
          "published_at": "2026-09-16T03:44:28.568777Z",
          "score": 100,
          "model": "gpt-6-1-sol",
          "runner": "benchmark_agent",
          "config": "openai-gpt-6-1-sol-xhigh-provider-adapter",
          "verified_publish": true,
          "levels_completed": 6,
          "actions": 149,
          "resets": 0,
          "state": "WIN",
          "completed": true,
          "number_of_levels": 0,
          "level_scores": [
            115,
            115,
            115,
            115,
            115,
            115
          ],
          "level_actions": [
            23,
            27,
            26,
            21,
            25,
            27
          ],
          "level_baseline_actions": [
            54,
            58,
            40,
            45,
            71,
            146
          ]
        },
        {
          "session_id": "2003b652-5031-4573-8bb7-eb8a806318a1",
          "game_id": "tr87",
          "game_version": "cd924810",
          "published_at": "2026-09-16T03:24:49.332429Z",
          "score": 0,
          "model": "gpt-6-luna",
          "runner": "benchmark_agent",
          "config": "openai-gpt-6-luna-max-provider-adapter",
          "verified_publish": true,
          "levels_completed": 0,
          "actions": 270,
          "resets": 3,
          "state": "GAME_OVER",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            270,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            54,
            58,
            40,
            45,
            71,
            146
          ]
        },
        {
          "session_id": "2429d807-8442-42be-86ce-5cbe9b0ccd77",
          "game_id": "tr87",
          "game_version": "cd924810",
          "published_at": "2026-09-16T01:24:41.697301Z",
          "score": 100,
          "model": "gpt-6-1-sol",
          "runner": "benchmark_agent",
          "config": "openai-gpt-6-1-sol-max-provider-adapter",
          "verified_publish": true,
          "levels_completed": 6,
          "actions": 152,
          "resets": 0,
          "state": "WIN",
          "completed": true,
          "number_of_levels": 0,
          "level_scores": [
            115,
            115,
            115,
            115,
            115,
            115
          ],
          "level_actions": [
            23,
            33,
            26,
            21,
            25,
            24
          ],
          "level_baseline_actions": [
            54,
            58,
            40,
            45,
            71,
            146
          ]
        },
        {
          "session_id": "4bc60a69-c0fe-42e7-bec7-61bc5115e56b",
          "game_id": "tr87",
          "game_version": "cd924810",
          "published_at": "2026-09-15T22:48:03.817289Z",
          "score": 100,
          "model": "gpt-6-1-sol",
          "runner": "benchmark_agent",
          "config": "openai-gpt-6-1-sol-max",
          "verified_publish": true,
          "levels_completed": 6,
          "actions": 171,
          "resets": 0,
          "state": "WIN",
          "completed": true,
          "number_of_levels": 0,
          "level_scores": [
            115,
            115,
            115,
            115,
            115,
            115
          ],
          "level_actions": [
            25,
            34,
            26,
            24,
            28,
            34
          ],
          "level_baseline_actions": [
            54,
            58,
            40,
            45,
            71,
            146
          ]
        },
        {
          "session_id": "6c879e58-b5b4-4792-9058-89bda88cda4b",
          "game_id": "tr87",
          "game_version": "cd924810",
          "published_at": "2026-09-12T14:05:27.069528Z",
          "score": 0,
          "model": "gemini-3.8-flash",
          "runner": "benchmark_agent",
          "config": "google-gemini-3-8-flash-low",
          "verified_publish": true,
          "levels_completed": 0,
          "actions": 270,
          "resets": 2,
          "state": "GAME_OVER",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            270,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            54,
            58,
            40,
            45,
            71,
            146
          ]
        },
        {
          "session_id": "2c0d0443-020f-4039-900b-544b439a869d",
          "game_id": "tr87",
          "game_version": "cd924810",
          "published_at": "2026-09-11T01:27:04.771337Z",
          "score": 0,
          "model": "gemini-3.8-flash",
          "runner": "benchmark_agent",
          "config": "google-gemini-3-8-flash-medium",
          "verified_publish": true,
          "levels_completed": 0,
          "actions": 270,
          "resets": 2,
          "state": "NOT_FINISHED",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            270,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            54,
            58,
            40,
            45,
            71,
            146
          ]
        },
        {
          "session_id": "9b851c36-3987-4b21-855b-e8c25d71c757",
          "game_id": "tr87",
          "game_version": "cd924810",
          "published_at": "2026-09-10T23:07:28.638123Z",
          "score": 0,
          "model": "gemini-3.8-flash",
          "runner": "benchmark_agent",
          "config": "google-gemini-3-8-flash-high",
          "verified_publish": true,
          "levels_completed": 0,
          "actions": 270,
          "resets": 6,
          "state": "NOT_FINISHED",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            270,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            54,
            58,
            40,
            45,
            71,
            146
          ]
        },
        {
          "session_id": "e0c28dfd-b5f2-4bc9-bffb-002c4b4cd4fb",
          "game_id": "tr87",
          "game_version": "cd924810",
          "published_at": "2026-08-11T12:26:46.245963Z",
          "score": 0,
          "model": "grok-4.6",
          "runner": "benchmark_agent",
          "config": "xai-grok-4-6-xhigh",
          "verified_publish": true,
          "levels_completed": 0,
          "actions": 270,
          "resets": 25,
          "state": "NOT_FINISHED",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            270,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            54,
            58,
            40,
            45,
            71,
            146
          ]
        },
        {
          "session_id": "0c60799e-d6ac-498b-aa70-29981e5ff094",
          "game_id": "tr87",
          "game_version": "cd924810",
          "published_at": "2026-07-21T11:57:54.889737Z",
          "score": 47.61904761904761,
          "model": "claude-opus-5",
          "runner": "benchmark_agent",
          "config": "anthropic-claude-opus-5-high",
          "verified_publish": true,
          "levels_completed": 4,
          "actions": 488,
          "resets": 2,
          "state": "GAME_OVER",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            115,
            115,
            105.19395134779748,
            115,
            0,
            0
          ],
          "level_actions": [
            35,
            32,
            39,
            27,
            355,
            0
          ],
          "level_baseline_actions": [
            54,
            58,
            40,
            45,
            71,
            146
          ]
        },
        {
          "session_id": "d46f7668-bac0-4193-8c75-25c3321bb037",
          "game_id": "tr87",
          "game_version": "cd924810",
          "published_at": "2026-06-24T11:01:49.491306Z",
          "score": 0,
          "model": "gpt-5-6-luna",
          "runner": "benchmark_agent",
          "config": "openai-gpt-5-6-luna-low",
          "verified_publish": true,
          "levels_completed": 0,
          "actions": 270,
          "resets": 18,
          "state": "NOT_FINISHED",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            270,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            54,
            58,
            40,
            45,
            71,
            146
          ]
        },
        {
          "session_id": "834dd916-70a1-4e87-bdc7-e3ba097d024e",
          "game_id": "tr87",
          "game_version": "cd924810",
          "published_at": "2026-06-24T10:49:40.911194Z",
          "score": 0,
          "model": "gpt-5-6-luna",
          "runner": "benchmark_agent",
          "config": "openai-gpt-5-6-luna-medium",
          "verified_publish": true,
          "levels_completed": 0,
          "actions": 270,
          "resets": 7,
          "state": "NOT_FINISHED",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            270,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            54,
            58,
            40,
            45,
            71,
            146
          ]
        },
        {
          "session_id": "b63d76e5-f316-4176-87a9-111f7db019ff",
          "game_id": "tr87",
          "game_version": "cd924810",
          "published_at": "2026-06-24T06:46:33.970011Z",
          "score": 0,
          "model": "gpt-5-6-luna",
          "runner": "benchmark_agent",
          "config": "openai-gpt-5-6-luna-high",
          "verified_publish": true,
          "levels_completed": 0,
          "actions": 270,
          "resets": 42,
          "state": "NOT_FINISHED",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            270,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            54,
            58,
            40,
            45,
            71,
            146
          ]
        },
        {
          "session_id": "9c6ae0b0-fa6f-47df-9fa5-a3d191d96957",
          "game_id": "tr87",
          "game_version": "cd924810",
          "published_at": "2026-06-24T05:41:38.892809Z",
          "score": 0,
          "model": "gpt-5-6-luna",
          "runner": "benchmark_agent",
          "config": "openai-gpt-5-6-luna-xhigh",
          "verified_publish": true,
          "levels_completed": 0,
          "actions": 270,
          "resets": 9,
          "state": "NOT_FINISHED",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            270,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            54,
            58,
            40,
            45,
            71,
            146
          ]
        },
        {
          "session_id": "a974eeed-595a-43c5-abda-750766a9f856",
          "game_id": "tr87",
          "game_version": "cd924810",
          "published_at": "2026-06-23T03:00:03.687975Z",
          "score": 0,
          "model": "gpt-5-6-luna",
          "runner": "benchmark_agent",
          "config": "openai-gpt-5-6-luna-max",
          "verified_publish": true,
          "levels_completed": 0,
          "actions": 64,
          "resets": 5,
          "state": "NOT_FINISHED",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            64,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            54,
            58,
            40,
            45,
            71,
            146
          ]
        },
        {
          "session_id": "ef2c4b36-8d3a-4b49-b266-39eb38d0e3b4",
          "game_id": "tr87",
          "game_version": "cd924810",
          "published_at": "2026-06-20T18:18:35.407123Z",
          "score": 0,
          "model": "gpt-5-6-terra",
          "runner": "benchmark_agent",
          "config": "openai-gpt-5-6-terra-xhigh",
          "verified_publish": true,
          "levels_completed": 0,
          "actions": 270,
          "resets": 42,
          "state": "NOT_FINISHED",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            270,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            54,
            58,
            40,
            45,
            71,
            146
          ]
        },
        {
          "session_id": "8c51fb2e-6c8f-4685-8aec-e01cd15491e7",
          "game_id": "tr87",
          "game_version": "cd924810",
          "published_at": "2026-06-20T12:01:30.293573Z",
          "score": 0,
          "model": "gpt-5-6-terra",
          "runner": "benchmark_agent",
          "config": "openai-gpt-5-6-terra-low",
          "verified_publish": true,
          "levels_completed": 0,
          "actions": 59,
          "resets": 0,
          "state": "NOT_FINISHED",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            59,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            54,
            58,
            40,
            45,
            71,
            146
          ]
        },
        {
          "session_id": "d13e070b-f3d0-4316-a2fe-3157f160a491",
          "game_id": "tr87",
          "game_version": "cd924810",
          "published_at": "2026-06-20T09:51:19.14197Z",
          "score": 0,
          "model": "gpt-5-6-terra",
          "runner": "benchmark_agent",
          "config": "openai-gpt-5-6-terra-high",
          "verified_publish": true,
          "levels_completed": 0,
          "actions": 270,
          "resets": 3,
          "state": "GAME_OVER",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            270,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            54,
            58,
            40,
            45,
            71,
            146
          ]
        },
        {
          "session_id": "a8896490-b14f-46c2-afab-994c05e534c1",
          "game_id": "tr87",
          "game_version": "cd924810",
          "published_at": "2026-06-20T09:42:06.234296Z",
          "score": 0,
          "model": "gpt-5-6-terra",
          "runner": "benchmark_agent",
          "config": "openai-gpt-5-6-terra-medium",
          "verified_publish": true,
          "levels_completed": 0,
          "actions": 270,
          "resets": 2,
          "state": "GAME_OVER",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            270,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            54,
            58,
            40,
            45,
            71,
            146
          ]
        },
        {
          "session_id": "35ac4cb2-b09e-4c0d-8886-d772acc1e05b",
          "game_id": "tr87",
          "game_version": "cd924810",
          "published_at": "2026-06-19T12:09:01.939597Z",
          "score": 0,
          "model": "gpt-5-6-terra",
          "runner": "benchmark_agent",
          "config": "openai-gpt-5-6-terra-max",
          "verified_publish": true,
          "levels_completed": 0,
          "actions": 270,
          "resets": 53,
          "state": "NOT_FINISHED",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            270,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            54,
            58,
            40,
            45,
            71,
            146
          ]
        },
        {
          "session_id": "c3f272ff-42e2-437a-bf4b-0f725c2ffcc9",
          "game_id": "tr87",
          "game_version": "cd924810",
          "published_at": "2026-06-14T01:09:43.914524Z",
          "score": 0.2257692879441058,
          "model": "gpt-5-6-sol",
          "runner": "benchmark_agent",
          "config": "openai-gpt-5-6-sol-max",
          "verified_publish": true,
          "levels_completed": 1,
          "actions": 344,
          "resets": 7,
          "state": "NOT_FINISHED",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            4.741155046826222,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            248,
            96,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            54,
            58,
            40,
            45,
            71,
            146
          ]
        },
        {
          "session_id": "e4c5835a-5eff-4b5b-9c46-82f3d8730e33",
          "game_id": "tr87",
          "game_version": "cd924810",
          "published_at": "2026-06-12T11:13:55.482634Z",
          "score": 0,
          "model": "gpt-5-6-sol",
          "runner": "benchmark_agent",
          "config": "openai-gpt-5-6-sol-high",
          "verified_publish": true,
          "levels_completed": 0,
          "actions": 270,
          "resets": 8,
          "state": "NOT_FINISHED",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            270,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            54,
            58,
            40,
            45,
            71,
            146
          ]
        },
        {
          "session_id": "0df5c49c-ef41-4910-8dc2-e0f9e9fa69d2",
          "game_id": "tr87",
          "game_version": "cd924810",
          "published_at": "2026-06-12T00:26:54.407463Z",
          "score": 0,
          "model": "gpt-5-6-sol",
          "runner": "benchmark_agent",
          "config": "openai-gpt-5-6-sol-xhigh",
          "verified_publish": true,
          "levels_completed": 0,
          "actions": 77,
          "resets": 4,
          "state": "NOT_FINISHED",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            77,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            54,
            58,
            40,
            45,
            71,
            146
          ]
        },
        {
          "session_id": "34b14674-134b-4a41-b708-71858fb7b91d",
          "game_id": "tr87",
          "game_version": "cd924810",
          "published_at": "2026-06-10T11:37:39.171355Z",
          "score": 0,
          "model": "gpt-5-6-sol",
          "runner": "benchmark_agent",
          "config": "openai-gpt-5-6-sol-medium",
          "verified_publish": true,
          "levels_completed": 0,
          "actions": 270,
          "resets": 4,
          "state": "NOT_FINISHED",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            270,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            54,
            58,
            40,
            45,
            71,
            146
          ]
        },
        {
          "session_id": "15dc1239-23de-4fcd-95fe-d47bab67d854",
          "game_id": "tr87",
          "game_version": "cd924810",
          "published_at": "2026-06-09T22:54:09.153895Z",
          "score": 0,
          "model": "gpt-5-6-sol",
          "runner": "benchmark_agent",
          "config": "openai-gpt-5-6-sol-low",
          "verified_publish": true,
          "levels_completed": 0,
          "actions": 270,
          "resets": 2,
          "state": "GAME_OVER",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            270,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            54,
            58,
            40,
            45,
            71,
            146
          ]
        },
        {
          "session_id": "c3b77031-b814-464e-b636-d9663ffd12a2",
          "game_id": "tr87",
          "game_version": "cd924810",
          "published_at": "2026-05-29T23:05:48.858114Z",
          "score": 0,
          "model": "claude-opus-4-8",
          "runner": "benchmark_agent",
          "config": "anthropic-opus-4-8-high",
          "verified_publish": true,
          "levels_completed": 0,
          "actions": 236,
          "resets": 3,
          "state": "NOT_FINISHED",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            236,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            54,
            58,
            40,
            45,
            71,
            146
          ]
        },
        {
          "session_id": "5a10b18d-8e04-496b-b6e9-b4382d36c7fb",
          "game_id": "tr87",
          "game_version": "cd924810",
          "published_at": "2026-05-01T16:12:58.016304Z",
          "score": 0,
          "model": "claude-opus-4-7",
          "runner": "benchmark_agent",
          "config": "anthropic-opus-4-7-high",
          "verified_publish": true,
          "levels_completed": 0,
          "actions": 270,
          "resets": 14,
          "state": "NOT_FINISHED",
          "completed": false,
          "number_of_levels": 6,
          "level_scores": [
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            270,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            54,
            58,
            40,
            45,
            71,
            146
          ]
        },
        {
          "session_id": "5895b428-17b9-4257-bd27-68481d5f6800",
          "game_id": "tr87",
          "game_version": "cd924810",
          "published_at": "2026-04-30T23:23:22.08897Z",
          "score": 0,
          "model": "gpt-5.5-2026-04-23",
          "runner": "benchmark_agent",
          "config": "openai-gpt-5-5-2026-04-23-high",
          "verified_publish": true,
          "levels_completed": 0,
          "actions": 270,
          "resets": 3,
          "state": "NOT_FINISHED",
          "completed": false,
          "number_of_levels": 6,
          "level_scores": [
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            270,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            54,
            58,
            40,
            45,
            71,
            146
          ]
        },
        {
          "session_id": "56c2a94a-1835-498b-bc37-2a07a5f14724",
          "game_id": "tr87",
          "game_version": "cd924810",
          "published_at": "2026-04-14T17:43:51.92172Z",
          "score": 0,
          "model": "grok-4.20-beta-0309-reasoning",
          "runner": "benchmark_agent",
          "config": "xai-grok-4-20-beta-0309-reasoning",
          "verified_publish": true,
          "levels_completed": 0,
          "actions": 68,
          "resets": 6,
          "state": "NOT_FINISHED",
          "completed": false,
          "number_of_levels": 6,
          "level_scores": [
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            68,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            54,
            58,
            40,
            45,
            71,
            146
          ]
        },
        {
          "session_id": "8a39f8d1-dadd-4e56-80e8-54fc138d2320",
          "game_id": "tr87",
          "game_version": "cd924810",
          "published_at": "2026-04-14T17:43:51.661913Z",
          "score": 0,
          "model": "gpt-5.4-2026-03-05",
          "runner": "benchmark_agent",
          "config": "openai-gpt-5-4-2026-03-05-high",
          "verified_publish": true,
          "levels_completed": 0,
          "actions": 270,
          "resets": 5,
          "state": "NOT_FINISHED",
          "completed": false,
          "number_of_levels": 6,
          "level_scores": [
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            270,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            54,
            58,
            40,
            45,
            71,
            146
          ]
        },
        {
          "session_id": "c7e6f80e-3cac-49fd-b690-7b54d006a769",
          "game_id": "tr87",
          "game_version": "cd924810",
          "published_at": "2026-04-14T17:43:51.208465Z",
          "score": 0,
          "model": "gemini-3.1-pro-preview",
          "runner": "benchmark_agent",
          "config": "google-gemini-3-1-pro-preview",
          "verified_publish": true,
          "levels_completed": 0,
          "actions": 199,
          "resets": 4,
          "state": "NOT_FINISHED",
          "completed": false,
          "number_of_levels": 6,
          "level_scores": [
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            199,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            54,
            58,
            40,
            45,
            71,
            146
          ]
        },
        {
          "session_id": "2f94f0bf-5c98-45f5-abf1-b980a82ea609",
          "game_id": "tr87",
          "game_version": "cd924810",
          "published_at": "2026-04-14T17:43:50.947217Z",
          "score": 0,
          "model": "claude-opus-4-6",
          "runner": "benchmark_agent",
          "config": "anthropic-opus-4-6-max-effort",
          "verified_publish": true,
          "levels_completed": 0,
          "actions": 270,
          "resets": 6,
          "state": "GAME_OVER",
          "completed": false,
          "number_of_levels": 6,
          "level_scores": [
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            270,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            54,
            58,
            40,
            45,
            71,
            146
          ]
        }
      ]
    },
    {
      "game": "sc25",
      "source": "https://arcprize.org/tasks/sc25",
      "runs": [
        {
          "session_id": "be04e0dd-d984-4334-ab6d-b2587a6d83aa",
          "game_id": "sc25",
          "game_version": "635fd71a",
          "published_at": "2026-09-25T12:12:47.236996Z",
          "score": 0,
          "model": "gpt-6-sol",
          "runner": "benchmark_agent",
          "config": "openai-gpt-6-sol-none-provider-adapter",
          "verified_publish": true,
          "levels_completed": 0,
          "actions": 172,
          "resets": 4,
          "state": "GAME_OVER",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            172,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            36,
            6,
            32,
            83,
            143,
            50
          ]
        },
        {
          "session_id": "dd9b5ec8-1a47-4205-90fd-9297d51afad3",
          "game_id": "sc25",
          "game_version": "635fd71a",
          "published_at": "2026-09-25T11:51:05.086395Z",
          "score": 19.035351536736577,
          "model": "gpt-6-sol",
          "runner": "benchmark_agent",
          "config": "openai-gpt-6-sol-max",
          "verified_publish": true,
          "levels_completed": 3,
          "actions": 488,
          "resets": 10,
          "state": "GAME_OVER",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            115,
            36,
            70.9141274238227,
            0,
            0,
            0
          ],
          "level_actions": [
            25,
            10,
            38,
            415,
            0,
            0
          ],
          "level_baseline_actions": [
            36,
            6,
            32,
            83,
            143,
            50
          ]
        },
        {
          "session_id": "5eb1f092-84f4-4fe6-80aa-728de727a649",
          "game_id": "sc25",
          "game_version": "635fd71a",
          "published_at": "2026-09-25T06:18:15.177222Z",
          "score": 0,
          "model": "gpt-6-sol",
          "runner": "benchmark_agent",
          "config": "openai-gpt-6-sol-none",
          "verified_publish": true,
          "levels_completed": 0,
          "actions": 180,
          "resets": 2,
          "state": "GAME_OVER",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            180,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            36,
            6,
            32,
            83,
            143,
            50
          ]
        },
        {
          "session_id": "de1c836a-2296-4899-a742-9f88fe1db8a7",
          "game_id": "sc25",
          "game_version": "635fd71a",
          "published_at": "2026-09-25T06:17:48.891889Z",
          "score": 14.140102321920505,
          "model": "gpt-6-sol",
          "runner": "benchmark_agent",
          "config": "openai-gpt-6-sol-low-provider-adapter",
          "verified_publish": true,
          "levels_completed": 2,
          "actions": 209,
          "resets": 8,
          "state": "GAME_OVER",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            66.94214876033058,
            115,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            44,
            5,
            160,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            36,
            6,
            32,
            83,
            143,
            50
          ]
        },
        {
          "session_id": "d1d0a4bc-990c-4b5f-9f94-e0767c804736",
          "game_id": "sc25",
          "game_version": "635fd71a",
          "published_at": "2026-09-25T06:17:47.996925Z",
          "score": 15.547225317271687,
          "model": "gpt-6-sol",
          "runner": "benchmark_agent",
          "config": "openai-gpt-6-sol-medium-provider-adapter",
          "verified_publish": true,
          "levels_completed": 3,
          "actions": 525,
          "resets": 10,
          "state": "GAME_OVER",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            38.525564803805,
            11.11111111111111,
            88.58131487889274,
            0,
            0,
            0
          ],
          "level_actions": [
            58,
            18,
            34,
            415,
            0,
            0
          ],
          "level_baseline_actions": [
            36,
            6,
            32,
            83,
            143,
            50
          ]
        },
        {
          "session_id": "938f24c9-7b4a-42d6-b782-4101eae900f9",
          "game_id": "sc25",
          "game_version": "635fd71a",
          "published_at": "2026-09-25T05:57:30.988869Z",
          "score": 0,
          "model": "gpt-6-sol",
          "runner": "benchmark_agent",
          "config": "openai-gpt-6-sol-low",
          "verified_publish": true,
          "levels_completed": 0,
          "actions": 180,
          "resets": 3,
          "state": "GAME_OVER",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            180,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            36,
            6,
            32,
            83,
            143,
            50
          ]
        },
        {
          "session_id": "e233778c-aa5b-4e46-83d8-e6efd2e99fc5",
          "game_id": "sc25",
          "game_version": "635fd71a",
          "published_at": "2026-09-25T05:48:53.284297Z",
          "score": 8.441881674094974,
          "model": "gpt-6-sol",
          "runner": "benchmark_agent",
          "config": "openai-gpt-6-sol-high-provider-adapter",
          "verified_publish": true,
          "levels_completed": 3,
          "actions": 656,
          "resets": 17,
          "state": "GAME_OVER",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            14.667270258035309,
            73.46938775510203,
            5.224489795918367,
            0,
            0,
            0
          ],
          "level_actions": [
            94,
            7,
            140,
            415,
            0,
            0
          ],
          "level_baseline_actions": [
            36,
            6,
            32,
            83,
            143,
            50
          ]
        },
        {
          "session_id": "6d7311ba-5f80-4247-8760-2343709338f3",
          "game_id": "sc25",
          "game_version": "635fd71a",
          "published_at": "2026-09-25T05:19:51.500474Z",
          "score": 0.22395952139020797,
          "model": "gpt-6-sol",
          "runner": "benchmark_agent",
          "config": "openai-gpt-6-sol-medium",
          "verified_publish": true,
          "levels_completed": 1,
          "actions": 196,
          "resets": 2,
          "state": "GAME_OVER",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            4.703149949194367,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            166,
            30,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            36,
            6,
            32,
            83,
            143,
            50
          ]
        },
        {
          "session_id": "695e873e-0eda-4064-8e13-77bd1f19a3c1",
          "game_id": "sc25",
          "game_version": "635fd71a",
          "published_at": "2026-09-25T04:55:12.249668Z",
          "score": 21.518922397300777,
          "model": "gpt-6-sol",
          "runner": "benchmark_agent",
          "config": "openai-gpt-6-sol-xhigh-provider-adapter",
          "verified_publish": true,
          "levels_completed": 3,
          "actions": 493,
          "resets": 11,
          "state": "GAME_OVER",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            115,
            56.25,
            74.7991234477721,
            0,
            0,
            0
          ],
          "level_actions": [
            33,
            8,
            37,
            415,
            0,
            0
          ],
          "level_baseline_actions": [
            36,
            6,
            32,
            83,
            143,
            50
          ]
        },
        {
          "session_id": "a65d0d39-785d-4adf-a829-b048ef7205e4",
          "game_id": "sc25",
          "game_version": "635fd71a",
          "published_at": "2026-09-25T04:49:29.610285Z",
          "score": 6.396807088848611,
          "model": "gpt-6-sol",
          "runner": "benchmark_agent",
          "config": "openai-gpt-6-sol-high",
          "verified_publish": true,
          "levels_completed": 2,
          "actions": 212,
          "resets": 3,
          "state": "GAME_OVER",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            112.11072664359862,
            11.11111111111111,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            34,
            18,
            160,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            36,
            6,
            32,
            83,
            143,
            50
          ]
        },
        {
          "session_id": "e18ccf34-8bcd-431b-b6a3-cabf53dd2f4b",
          "game_id": "sc25",
          "game_version": "635fd71a",
          "published_at": "2026-09-25T04:09:42.659685Z",
          "score": 2.6785714285714284,
          "model": "gpt-6-sol",
          "runner": "benchmark_agent",
          "config": "openai-gpt-6-sol-xhigh",
          "verified_publish": true,
          "levels_completed": 1,
          "actions": 78,
          "resets": 1,
          "state": "GAME_OVER",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            56.25,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            48,
            30,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            36,
            6,
            32,
            83,
            143,
            50
          ]
        },
        {
          "session_id": "0a6d4bb7-e34b-427e-950d-93891fd0ba6a",
          "game_id": "sc25",
          "game_version": "635fd71a",
          "published_at": "2026-09-21T14:02:37.326318Z",
          "score": 35.954707728769705,
          "model": "gemini-3.8-flash",
          "runner": "benchmark_agent",
          "config": "google-gemini-3-8-flash-low-provider-adapter",
          "verified_publish": true,
          "levels_completed": 5,
          "actions": 845,
          "resets": 14,
          "state": "GAME_OVER",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            6.079939951210358,
            56.25,
            4.6749452154857565,
            11.861021676624024,
            115,
            0
          ],
          "level_actions": [
            146,
            8,
            148,
            241,
            52,
            250
          ],
          "level_baseline_actions": [
            36,
            6,
            32,
            83,
            143,
            50
          ]
        },
        {
          "session_id": "0c34c8cd-d863-41df-b624-28a26b911fd4",
          "game_id": "sc25",
          "game_version": "635fd71a",
          "published_at": "2026-09-20T11:02:48.955058Z",
          "score": 47.61904761904761,
          "model": "gemini-3.8-flash",
          "runner": "benchmark_agent",
          "config": "google-gemini-3-8-flash-medium-provider-adapter",
          "verified_publish": true,
          "levels_completed": 4,
          "actions": 799,
          "resets": 10,
          "state": "GAME_OVER",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            115,
            73.46938775510203,
            115,
            115,
            0,
            0
          ],
          "level_actions": [
            20,
            7,
            29,
            28,
            715,
            0
          ],
          "level_baseline_actions": [
            36,
            6,
            32,
            83,
            143,
            50
          ]
        },
        {
          "session_id": "b11374ef-ff59-4ad8-8e22-52b6006b5da3",
          "game_id": "sc25",
          "game_version": "635fd71a",
          "published_at": "2026-09-17T22:50:28.317768Z",
          "score": 71.42857142857143,
          "model": "gemini-3.8-flash",
          "runner": "benchmark_agent",
          "config": "google-gemini-3-8-flash-high-provider-adapter",
          "verified_publish": true,
          "levels_completed": 5,
          "actions": 394,
          "resets": 4,
          "state": "GAME_OVER",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            115,
            115,
            115,
            115,
            115,
            0
          ],
          "level_actions": [
            24,
            5,
            18,
            32,
            65,
            250
          ],
          "level_baseline_actions": [
            36,
            6,
            32,
            83,
            143,
            50
          ]
        },
        {
          "session_id": "19f91136-5c42-47e2-ad03-8406a443e380",
          "game_id": "sc25",
          "game_version": "635fd71a",
          "published_at": "2026-09-16T20:27:50.56902Z",
          "score": 0,
          "model": "gpt-6-luna",
          "runner": "benchmark_agent",
          "config": "openai-gpt-6-luna-low",
          "verified_publish": true,
          "levels_completed": 0,
          "actions": 180,
          "resets": 3,
          "state": "GAME_OVER",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            180,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            36,
            6,
            32,
            83,
            143,
            50
          ]
        },
        {
          "session_id": "b8bffc62-77a5-486c-a06d-ba2ffa7982a8",
          "game_id": "sc25",
          "game_version": "635fd71a",
          "published_at": "2026-09-16T20:27:06.138667Z",
          "score": 0,
          "model": "gpt-6-luna",
          "runner": "benchmark_agent",
          "config": "openai-gpt-6-luna-none-provider-adapter",
          "verified_publish": true,
          "levels_completed": 0,
          "actions": 180,
          "resets": 1,
          "state": "GAME_OVER",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            180,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            36,
            6,
            32,
            83,
            143,
            50
          ]
        },
        {
          "session_id": "b9b9b754-a321-4916-8713-0d0c0d17d5f9",
          "game_id": "sc25",
          "game_version": "635fd71a",
          "published_at": "2026-09-16T19:32:18.74862Z",
          "score": 0,
          "model": "gpt-6-luna",
          "runner": "benchmark_agent",
          "config": "openai-gpt-6-luna-low-provider-adapter",
          "verified_publish": true,
          "levels_completed": 0,
          "actions": 180,
          "resets": 2,
          "state": "GAME_OVER",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            180,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            36,
            6,
            32,
            83,
            143,
            50
          ]
        },
        {
          "session_id": "1e9469da-291f-467a-ac1f-a29b1295bb5c",
          "game_id": "sc25",
          "game_version": "635fd71a",
          "published_at": "2026-09-16T19:26:52.133416Z",
          "score": 100,
          "model": "gpt-6-1-sol",
          "runner": "benchmark_agent",
          "config": "openai-gpt-6-1-sol-high-provider-adapter",
          "verified_publish": true,
          "levels_completed": 6,
          "actions": 176,
          "resets": 0,
          "state": "WIN",
          "completed": true,
          "number_of_levels": 0,
          "level_scores": [
            115,
            115,
            115,
            115,
            115,
            115
          ],
          "level_actions": [
            25,
            5,
            12,
            25,
            66,
            43
          ],
          "level_baseline_actions": [
            36,
            6,
            32,
            83,
            143,
            50
          ]
        },
        {
          "session_id": "d943f5b5-e354-41c6-9af0-f11f9ec650a1",
          "game_id": "sc25",
          "game_version": "635fd71a",
          "published_at": "2026-09-16T19:25:00.108516Z",
          "score": 0,
          "model": "gpt-6-luna",
          "runner": "benchmark_agent",
          "config": "openai-gpt-6-luna-none",
          "verified_publish": true,
          "levels_completed": 0,
          "actions": 180,
          "resets": 4,
          "state": "NOT_FINISHED",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            180,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            36,
            6,
            32,
            83,
            143,
            50
          ]
        },
        {
          "session_id": "efb52598-e596-45fd-9511-cc0a151bbf62",
          "game_id": "sc25",
          "game_version": "635fd71a",
          "published_at": "2026-09-16T17:58:26.104436Z",
          "score": 0,
          "model": "gpt-6-luna",
          "runner": "benchmark_agent",
          "config": "openai-gpt-6-luna-medium-provider-adapter",
          "verified_publish": true,
          "levels_completed": 0,
          "actions": 180,
          "resets": 2,
          "state": "GAME_OVER",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            180,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            36,
            6,
            32,
            83,
            143,
            50
          ]
        },
        {
          "session_id": "4c73afc8-3c03-42b8-be87-18a8503f1497",
          "game_id": "sc25",
          "game_version": "635fd71a",
          "published_at": "2026-09-16T16:55:34.916563Z",
          "score": 0,
          "model": "gpt-6-luna",
          "runner": "benchmark_agent",
          "config": "openai-gpt-6-luna-xhigh",
          "verified_publish": true,
          "levels_completed": 0,
          "actions": 180,
          "resets": 2,
          "state": "GAME_OVER",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            180,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            36,
            6,
            32,
            83,
            143,
            50
          ]
        },
        {
          "session_id": "2ce4d883-709b-4239-a96a-de113c5d80bd",
          "game_id": "sc25",
          "game_version": "635fd71a",
          "published_at": "2026-09-16T15:48:09.90275Z",
          "score": 1.0684606252473288,
          "model": "gpt-6-luna",
          "runner": "benchmark_agent",
          "config": "openai-gpt-6-luna-high-provider-adapter",
          "verified_publish": true,
          "levels_completed": 1,
          "actions": 106,
          "resets": 1,
          "state": "GAME_OVER",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            22.437673130193904,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            76,
            30,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            36,
            6,
            32,
            83,
            143,
            50
          ]
        },
        {
          "session_id": "3a8b3da8-c53c-4435-9782-5a13dc5fffd0",
          "game_id": "sc25",
          "game_version": "635fd71a",
          "published_at": "2026-09-16T14:43:55.554538Z",
          "score": 4.761904761904762,
          "model": "gpt-6-1-sol",
          "runner": "benchmark_agent",
          "config": "openai-gpt-6-1-sol-low",
          "verified_publish": true,
          "levels_completed": 1,
          "actions": 55,
          "resets": 1,
          "state": "GAME_OVER",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            115,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            25,
            30,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            36,
            6,
            32,
            83,
            143,
            50
          ]
        },
        {
          "session_id": "8b2487f4-50e9-4b8a-9f83-428fd54f9432",
          "game_id": "sc25",
          "game_version": "635fd71a",
          "published_at": "2026-09-16T09:06:00.255813Z",
          "score": 0,
          "model": "gpt-6-luna",
          "runner": "benchmark_agent",
          "config": "openai-gpt-6-luna-medium",
          "verified_publish": true,
          "levels_completed": 0,
          "actions": 180,
          "resets": 2,
          "state": "GAME_OVER",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            180,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            36,
            6,
            32,
            83,
            143,
            50
          ]
        },
        {
          "session_id": "9e078550-a87b-4784-b027-a6b5abefe85c",
          "game_id": "sc25",
          "game_version": "635fd71a",
          "published_at": "2026-09-16T09:00:22.297238Z",
          "score": 0,
          "model": "gpt-6-luna",
          "runner": "benchmark_agent",
          "config": "openai-gpt-6-luna-max",
          "verified_publish": true,
          "levels_completed": 0,
          "actions": 180,
          "resets": 4,
          "state": "GAME_OVER",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            180,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            36,
            6,
            32,
            83,
            143,
            50
          ]
        },
        {
          "session_id": "31e375d5-477c-4f79-9766-46d626f48867",
          "game_id": "sc25",
          "game_version": "635fd71a",
          "published_at": "2026-09-16T08:56:21.18169Z",
          "score": 0,
          "model": "gpt-6-luna",
          "runner": "benchmark_agent",
          "config": "openai-gpt-6-luna-high",
          "verified_publish": true,
          "levels_completed": 0,
          "actions": 180,
          "resets": 2,
          "state": "GAME_OVER",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            180,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            36,
            6,
            32,
            83,
            143,
            50
          ]
        },
        {
          "session_id": "66b345cc-dcfb-4f3a-b346-ae1d751f4649",
          "game_id": "sc25",
          "game_version": "635fd71a",
          "published_at": "2026-09-16T08:03:04.40128Z",
          "score": 0,
          "model": "gpt-6-luna",
          "runner": "benchmark_agent",
          "config": "openai-gpt-6-luna-xhigh-provider-adapter",
          "verified_publish": true,
          "levels_completed": 0,
          "actions": 180,
          "resets": 3,
          "state": "GAME_OVER",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            180,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            36,
            6,
            32,
            83,
            143,
            50
          ]
        },
        {
          "session_id": "36abb09f-742c-4b11-93e3-2e2899602856",
          "game_id": "sc25",
          "game_version": "635fd71a",
          "published_at": "2026-09-16T07:20:59.701076Z",
          "score": 0,
          "model": "gpt-6-luna",
          "runner": "benchmark_agent",
          "config": "openai-gpt-6-luna-max-provider-adapter",
          "verified_publish": true,
          "levels_completed": 0,
          "actions": 180,
          "resets": 4,
          "state": "GAME_OVER",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            180,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            36,
            6,
            32,
            83,
            143,
            50
          ]
        },
        {
          "session_id": "bfb3f2dc-79da-4f60-80fd-6f7800bd301b",
          "game_id": "sc25",
          "game_version": "635fd71a",
          "published_at": "2026-09-16T05:50:09.93815Z",
          "score": 37.666749338624335,
          "model": "gpt-6-1-sol",
          "runner": "benchmark_agent",
          "config": "openai-gpt-6-1-sol-xhigh",
          "verified_publish": true,
          "levels_completed": 4,
          "actions": 869,
          "resets": 12,
          "state": "GAME_OVER",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            115,
            16.000000000000004,
            115,
            74.75043402777779,
            0,
            0
          ],
          "level_actions": [
            23,
            15,
            20,
            96,
            715,
            0
          ],
          "level_baseline_actions": [
            36,
            6,
            32,
            83,
            143,
            50
          ]
        },
        {
          "session_id": "32942ae1-9c25-4e95-bb2a-5a2fa1376592",
          "game_id": "sc25",
          "game_version": "635fd71a",
          "published_at": "2026-09-16T05:13:57.216099Z",
          "score": 42.02375121477162,
          "model": "gpt-6-1-sol",
          "runner": "benchmark_agent",
          "config": "openai-gpt-6-1-sol-max",
          "verified_publish": true,
          "levels_completed": 4,
          "actions": 852,
          "resets": 12,
          "state": "GAME_OVER",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            115,
            73.46938775510203,
            115,
            68.89,
            0,
            0
          ],
          "level_actions": [
            15,
            7,
            15,
            100,
            715,
            0
          ],
          "level_baseline_actions": [
            36,
            6,
            32,
            83,
            143,
            50
          ]
        },
        {
          "session_id": "2d9aa3f7-93a2-4c6c-be1a-2b710a1d8be2",
          "game_id": "sc25",
          "game_version": "635fd71a",
          "published_at": "2026-09-16T05:09:36.12806Z",
          "score": 6.662547371889933,
          "model": "gpt-6-1-sol",
          "runner": "benchmark_agent",
          "config": "openai-gpt-6-1-sol-medium",
          "verified_publish": true,
          "levels_completed": 2,
          "actions": 204,
          "resets": 2,
          "state": "GAME_OVER",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            115,
            12.456747404844293,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            27,
            17,
            160,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            36,
            6,
            32,
            83,
            143,
            50
          ]
        },
        {
          "session_id": "2b95a9e4-c153-42ed-acc3-c72e378edd62",
          "game_id": "sc25",
          "game_version": "635fd71a",
          "published_at": "2026-09-16T04:21:11.566715Z",
          "score": 23.24404761904762,
          "model": "gpt-6-1-sol",
          "runner": "benchmark_agent",
          "config": "openai-gpt-6-1-sol-high",
          "verified_publish": true,
          "levels_completed": 3,
          "actions": 467,
          "resets": 10,
          "state": "GAME_OVER",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            115,
            14.0625,
            115,
            0,
            0,
            0
          ],
          "level_actions": [
            17,
            16,
            19,
            415,
            0,
            0
          ],
          "level_baseline_actions": [
            36,
            6,
            32,
            83,
            143,
            50
          ]
        },
        {
          "session_id": "3381bfda-ba63-4162-82ca-c786edbf8de1",
          "game_id": "sc25",
          "game_version": "635fd71a",
          "published_at": "2026-09-16T03:25:02.38312Z",
          "score": 100,
          "model": "gpt-6-1-sol",
          "runner": "benchmark_agent",
          "config": "openai-gpt-6-1-sol-xhigh-provider-adapter",
          "verified_publish": true,
          "levels_completed": 6,
          "actions": 162,
          "resets": 0,
          "state": "WIN",
          "completed": true,
          "number_of_levels": 0,
          "level_scores": [
            115,
            100,
            115,
            115,
            115,
            115
          ],
          "level_actions": [
            17,
            6,
            17,
            25,
            61,
            36
          ],
          "level_baseline_actions": [
            36,
            6,
            32,
            83,
            143,
            50
          ]
        },
        {
          "session_id": "7c19fb2b-3065-4958-af41-c50ae586192e",
          "game_id": "sc25",
          "game_version": "635fd71a",
          "published_at": "2026-09-16T01:22:30.264789Z",
          "score": 100,
          "model": "gpt-6-1-sol",
          "runner": "benchmark_agent",
          "config": "openai-gpt-6-1-sol-max-provider-adapter",
          "verified_publish": true,
          "levels_completed": 6,
          "actions": 166,
          "resets": 0,
          "state": "WIN",
          "completed": true,
          "number_of_levels": 0,
          "level_scores": [
            115,
            115,
            115,
            115,
            115,
            115
          ],
          "level_actions": [
            31,
            5,
            12,
            25,
            57,
            36
          ],
          "level_baseline_actions": [
            36,
            6,
            32,
            83,
            143,
            50
          ]
        },
        {
          "session_id": "92fbc32f-d7f1-4d0e-8b46-22c89949b20e",
          "game_id": "sc25",
          "game_version": "635fd71a",
          "published_at": "2026-09-12T12:22:14.135562Z",
          "score": 19.590350968729346,
          "model": "gemini-3.8-flash",
          "runner": "benchmark_agent",
          "config": "google-gemini-3-8-flash-low",
          "verified_publish": true,
          "levels_completed": 3,
          "actions": 484,
          "resets": 10,
          "state": "GAME_OVER",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            115,
            36,
            74.7991234477721,
            0,
            0,
            0
          ],
          "level_actions": [
            22,
            10,
            37,
            415,
            0,
            0
          ],
          "level_baseline_actions": [
            36,
            6,
            32,
            83,
            143,
            50
          ]
        },
        {
          "session_id": "851d3ec3-60f6-47fa-a00f-0049c5793083",
          "game_id": "sc25",
          "game_version": "635fd71a",
          "published_at": "2026-09-09T19:11:46.121336Z",
          "score": 9.059673386340979,
          "model": "gemini-3.8-flash",
          "runner": "benchmark_agent",
          "config": "google-gemini-3-8-flash-medium",
          "verified_publish": true,
          "levels_completed": 3,
          "actions": 535,
          "resets": 11,
          "state": "GAME_OVER",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            115,
            21.301775147928996,
            10.883196939100861,
            0,
            0,
            0
          ],
          "level_actions": [
            31,
            13,
            97,
            394,
            0,
            0
          ],
          "level_baseline_actions": [
            36,
            6,
            32,
            83,
            143,
            50
          ]
        },
        {
          "session_id": "1830d41f-85f8-4f54-bd19-ae8b0f156880",
          "game_id": "sc25",
          "game_version": "635fd71a",
          "published_at": "2026-09-09T17:36:50.752107Z",
          "score": 24.738292011019283,
          "model": "gemini-3.8-flash",
          "runner": "benchmark_agent",
          "config": "google-gemini-3-8-flash-high",
          "verified_publish": true,
          "levels_completed": 3,
          "actions": 222,
          "resets": 4,
          "state": "GAME_OVER",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            115,
            29.752066115702476,
            115,
            0,
            0,
            0
          ],
          "level_actions": [
            32,
            11,
            28,
            151,
            0,
            0
          ],
          "level_baseline_actions": [
            36,
            6,
            32,
            83,
            143,
            50
          ]
        },
        {
          "session_id": "d140aa3c-6f7d-4f14-bd7f-e3f66d2c20b7",
          "game_id": "sc25",
          "game_version": "635fd71a",
          "published_at": "2026-08-11T06:37:13.714465Z",
          "score": 0,
          "model": "grok-4.6",
          "runner": "benchmark_agent",
          "config": "xai-grok-4-6-xhigh",
          "verified_publish": true,
          "levels_completed": 0,
          "actions": 180,
          "resets": 4,
          "state": "GAME_OVER",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            180,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            36,
            6,
            32,
            83,
            143,
            50
          ]
        },
        {
          "session_id": "c136ac3c-8cd3-49c7-b871-ef732fc4657a",
          "game_id": "sc25",
          "game_version": "635fd71a",
          "published_at": "2026-07-21T00:13:02.017507Z",
          "score": 0,
          "model": "claude-opus-5",
          "runner": "benchmark_agent",
          "config": "anthropic-claude-opus-5-high",
          "verified_publish": true,
          "levels_completed": 0,
          "actions": 180,
          "resets": 4,
          "state": "GAME_OVER",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            180,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            36,
            6,
            32,
            83,
            143,
            50
          ]
        },
        {
          "session_id": "9873aae9-c705-4b24-8da6-857e8cf1e14e",
          "game_id": "sc25",
          "game_version": "635fd71a",
          "published_at": "2026-06-24T10:49:06.268962Z",
          "score": 0,
          "model": "gpt-5-6-luna",
          "runner": "benchmark_agent",
          "config": "openai-gpt-5-6-luna-low",
          "verified_publish": true,
          "levels_completed": 0,
          "actions": 180,
          "resets": 9,
          "state": "GAME_OVER",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            180,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            36,
            6,
            32,
            83,
            143,
            50
          ]
        },
        {
          "session_id": "1cedc7a8-f563-4903-ae19-a14546486e91",
          "game_id": "sc25",
          "game_version": "635fd71a",
          "published_at": "2026-06-24T10:28:38.562242Z",
          "score": 0,
          "model": "gpt-5-6-luna",
          "runner": "benchmark_agent",
          "config": "openai-gpt-5-6-luna-medium",
          "verified_publish": true,
          "levels_completed": 0,
          "actions": 180,
          "resets": 13,
          "state": "NOT_FINISHED",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            180,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            36,
            6,
            32,
            83,
            143,
            50
          ]
        },
        {
          "session_id": "a9643cf8-c244-41ed-80fb-f2cce8253cf6",
          "game_id": "sc25",
          "game_version": "635fd71a",
          "published_at": "2026-06-24T03:20:25.720122Z",
          "score": 0,
          "model": "gpt-5-6-luna",
          "runner": "benchmark_agent",
          "config": "openai-gpt-5-6-luna-high",
          "verified_publish": true,
          "levels_completed": 0,
          "actions": 180,
          "resets": 10,
          "state": "GAME_OVER",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            180,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            36,
            6,
            32,
            83,
            143,
            50
          ]
        },
        {
          "session_id": "074e5a69-5f37-4c31-b0bc-404e0eaba8ba",
          "game_id": "sc25",
          "game_version": "635fd71a",
          "published_at": "2026-06-24T00:30:25.176689Z",
          "score": 0,
          "model": "gpt-5-6-luna",
          "runner": "benchmark_agent",
          "config": "openai-gpt-5-6-luna-xhigh",
          "verified_publish": true,
          "levels_completed": 0,
          "actions": 180,
          "resets": 10,
          "state": "NOT_FINISHED",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            180,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            36,
            6,
            32,
            83,
            143,
            50
          ]
        },
        {
          "session_id": "ffe5af9d-ba35-4838-ba58-3c65cfe937f5",
          "game_id": "sc25",
          "game_version": "635fd71a",
          "published_at": "2026-06-23T02:07:36.417898Z",
          "score": 0,
          "model": "gpt-5-6-luna",
          "runner": "benchmark_agent",
          "config": "openai-gpt-5-6-luna-max",
          "verified_publish": true,
          "levels_completed": 0,
          "actions": 180,
          "resets": 30,
          "state": "NOT_FINISHED",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            180,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            36,
            6,
            32,
            83,
            143,
            50
          ]
        },
        {
          "session_id": "6ed22cad-79cf-4153-827a-e6c86c2c3f2b",
          "game_id": "sc25",
          "game_version": "635fd71a",
          "published_at": "2026-06-20T11:59:12.801771Z",
          "score": 0,
          "model": "gpt-5-6-terra",
          "runner": "benchmark_agent",
          "config": "openai-gpt-5-6-terra-low",
          "verified_publish": true,
          "levels_completed": 0,
          "actions": 180,
          "resets": 3,
          "state": "GAME_OVER",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            180,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            36,
            6,
            32,
            83,
            143,
            50
          ]
        },
        {
          "session_id": "a5eec1b1-565d-4dee-9e8d-2cb60714f96d",
          "game_id": "sc25",
          "game_version": "635fd71a",
          "published_at": "2026-06-20T09:17:07.467825Z",
          "score": 0,
          "model": "gpt-5-6-terra",
          "runner": "benchmark_agent",
          "config": "openai-gpt-5-6-terra-medium",
          "verified_publish": true,
          "levels_completed": 0,
          "actions": 180,
          "resets": 4,
          "state": "GAME_OVER",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            180,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            36,
            6,
            32,
            83,
            143,
            50
          ]
        },
        {
          "session_id": "52991f95-fc04-442f-a041-eb3e50b78d57",
          "game_id": "sc25",
          "game_version": "635fd71a",
          "published_at": "2026-06-20T08:42:06.262033Z",
          "score": 0,
          "model": "gpt-5-6-terra",
          "runner": "benchmark_agent",
          "config": "openai-gpt-5-6-terra-high",
          "verified_publish": true,
          "levels_completed": 0,
          "actions": 180,
          "resets": 3,
          "state": "GAME_OVER",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            180,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            36,
            6,
            32,
            83,
            143,
            50
          ]
        },
        {
          "session_id": "d4f487e7-a22f-45fa-8db0-98b2df2ddc72",
          "game_id": "sc25",
          "game_version": "635fd71a",
          "published_at": "2026-06-20T01:19:20.635968Z",
          "score": 0.2706648204652678,
          "model": "gpt-5-6-terra",
          "runner": "benchmark_agent",
          "config": "openai-gpt-5-6-terra-xhigh",
          "verified_publish": true,
          "levels_completed": 1,
          "actions": 181,
          "resets": 3,
          "state": "GAME_OVER",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            5.683961229770624,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            151,
            30,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            36,
            6,
            32,
            83,
            143,
            50
          ]
        },
        {
          "session_id": "8a7379b2-7ae4-4836-904f-2b67bf662bb0",
          "game_id": "sc25",
          "game_version": "635fd71a",
          "published_at": "2026-06-18T20:48:29.028943Z",
          "score": 4.761904761904762,
          "model": "gpt-5-6-terra",
          "runner": "benchmark_agent",
          "config": "openai-gpt-5-6-terra-max",
          "verified_publish": true,
          "levels_completed": 1,
          "actions": 53,
          "resets": 0,
          "state": "GAME_OVER",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            115,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            23,
            30,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            36,
            6,
            32,
            83,
            143,
            50
          ]
        },
        {
          "session_id": "6693f6a6-8d4e-4a4e-95c0-e2ee7e1309bc",
          "game_id": "sc25",
          "game_version": "635fd71a",
          "published_at": "2026-06-13T10:10:18.447332Z",
          "score": 17.822840115975552,
          "model": "gpt-5-6-sol",
          "runner": "benchmark_agent",
          "config": "openai-gpt-5-6-sol-max",
          "verified_publish": true,
          "levels_completed": 3,
          "actions": 506,
          "resets": 12,
          "state": "GAME_OVER",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            64.00000000000001,
            29.752066115702476,
            83.59183673469387,
            0,
            0,
            0
          ],
          "level_actions": [
            45,
            11,
            35,
            415,
            0,
            0
          ],
          "level_baseline_actions": [
            36,
            6,
            32,
            83,
            143,
            50
          ]
        },
        {
          "session_id": "692f9d97-9326-44e8-95db-5fba10615c96",
          "game_id": "sc25",
          "game_version": "635fd71a",
          "published_at": "2026-06-11T15:34:22.602932Z",
          "score": 5.544973544973545,
          "model": "gpt-5-6-sol",
          "runner": "benchmark_agent",
          "config": "openai-gpt-5-6-sol-xhigh",
          "verified_publish": true,
          "levels_completed": 2,
          "actions": 224,
          "resets": 5,
          "state": "GAME_OVER",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            44.44444444444444,
            36,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            54,
            10,
            160,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            36,
            6,
            32,
            83,
            143,
            50
          ]
        },
        {
          "session_id": "4190a530-2589-4f2f-a035-8c47ddab5361",
          "game_id": "sc25",
          "game_version": "635fd71a",
          "published_at": "2026-06-10T12:37:41.806652Z",
          "score": 2.019935957491507,
          "model": "gpt-5-6-sol",
          "runner": "benchmark_agent",
          "config": "openai-gpt-5-6-sol-high",
          "verified_publish": true,
          "levels_completed": 2,
          "actions": 325,
          "resets": 7,
          "state": "GAME_OVER",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            5.683961229770624,
            18.36734693877551,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            151,
            14,
            160,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            36,
            6,
            32,
            83,
            143,
            50
          ]
        },
        {
          "session_id": "a3dd3183-47cb-4bca-983c-0efcca0ee25c",
          "game_id": "sc25",
          "game_version": "635fd71a",
          "published_at": "2026-06-10T08:07:37.734396Z",
          "score": 0,
          "model": "gpt-5-6-sol",
          "runner": "benchmark_agent",
          "config": "openai-gpt-5-6-sol-medium",
          "verified_publish": true,
          "levels_completed": 0,
          "actions": 180,
          "resets": 4,
          "state": "GAME_OVER",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            180,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            36,
            6,
            32,
            83,
            143,
            50
          ]
        },
        {
          "session_id": "8f202156-2ad6-4efd-a0fd-b396baa24b9a",
          "game_id": "sc25",
          "game_version": "635fd71a",
          "published_at": "2026-06-09T21:41:42.117431Z",
          "score": 0.2706648204652678,
          "model": "gpt-5-6-sol",
          "runner": "benchmark_agent",
          "config": "openai-gpt-5-6-sol-low",
          "verified_publish": true,
          "levels_completed": 1,
          "actions": 181,
          "resets": 3,
          "state": "GAME_OVER",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            5.683961229770624,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            151,
            30,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            36,
            6,
            32,
            83,
            143,
            50
          ]
        },
        {
          "session_id": "c50298eb-18e5-4049-bad7-dea8a616063c",
          "game_id": "sc25",
          "game_version": "635fd71a",
          "published_at": "2026-05-29T16:57:11.306636Z",
          "score": 0,
          "model": "claude-opus-4-8",
          "runner": "benchmark_agent",
          "config": "anthropic-opus-4-8-high",
          "verified_publish": true,
          "levels_completed": 0,
          "actions": 180,
          "resets": 4,
          "state": "GAME_OVER",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            180,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            36,
            6,
            32,
            83,
            143,
            50
          ]
        },
        {
          "session_id": "54ef2b19-3cf6-4027-816b-82a662540a72",
          "game_id": "sc25",
          "game_version": "635fd71a",
          "published_at": "2026-05-01T05:06:55.079467Z",
          "score": 0,
          "model": "claude-opus-4-7",
          "runner": "benchmark_agent",
          "config": "anthropic-opus-4-7-high",
          "verified_publish": true,
          "levels_completed": 0,
          "actions": 21,
          "resets": 2,
          "state": "NOT_FINISHED",
          "completed": false,
          "number_of_levels": 6,
          "level_scores": [
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            21,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            36,
            6,
            32,
            83,
            143,
            50
          ]
        },
        {
          "session_id": "34ca2c4c-0838-4f27-8948-a886eaf737f6",
          "game_id": "sc25",
          "game_version": "635fd71a",
          "published_at": "2026-04-30T23:24:14.778311Z",
          "score": 4.761904761904762,
          "model": "gpt-5.5-2026-04-23",
          "runner": "benchmark_agent",
          "config": "openai-gpt-5-5-2026-04-23-high",
          "verified_publish": true,
          "levels_completed": 1,
          "actions": 52,
          "resets": 1,
          "state": "GAME_OVER",
          "completed": false,
          "number_of_levels": 6,
          "level_scores": [
            115,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            22,
            30,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            36,
            6,
            32,
            83,
            143,
            50
          ]
        },
        {
          "session_id": "d5773752-a4ef-4336-9b4c-9b442968f1d3",
          "game_id": "sc25",
          "game_version": "635fd71a",
          "published_at": "2026-04-14T17:43:50.244143Z",
          "score": 0,
          "model": "grok-4.20-beta-0309-reasoning",
          "runner": "benchmark_agent",
          "config": "xai-grok-4-20-beta-0309-reasoning",
          "verified_publish": true,
          "levels_completed": 0,
          "actions": 180,
          "resets": 10,
          "state": "NOT_FINISHED",
          "completed": false,
          "number_of_levels": 6,
          "level_scores": [
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            180,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            36,
            6,
            32,
            83,
            143,
            50
          ]
        },
        {
          "session_id": "c74c9f49-8add-4e04-8c15-98f5ce887af7",
          "game_id": "sc25",
          "game_version": "635fd71a",
          "published_at": "2026-04-14T17:43:49.593855Z",
          "score": 0.2015160349854227,
          "model": "gpt-5.4-2026-03-05",
          "runner": "benchmark_agent",
          "config": "openai-gpt-5-4-2026-03-05-high",
          "verified_publish": true,
          "levels_completed": 1,
          "actions": 205,
          "resets": 3,
          "state": "GAME_OVER",
          "completed": false,
          "number_of_levels": 6,
          "level_scores": [
            4.231836734693877,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            175,
            30,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            36,
            6,
            32,
            83,
            143,
            50
          ]
        },
        {
          "session_id": "bd8e6a96-5b33-411a-871e-2f31cd7bb8b4",
          "game_id": "sc25",
          "game_version": "635fd71a",
          "published_at": "2026-04-14T17:43:48.095706Z",
          "score": 0,
          "model": "gemini-3.1-pro-preview",
          "runner": "benchmark_agent",
          "config": "google-gemini-3-1-pro-preview",
          "verified_publish": true,
          "levels_completed": 0,
          "actions": 147,
          "resets": 3,
          "state": "NOT_FINISHED",
          "completed": false,
          "number_of_levels": 6,
          "level_scores": [
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            147,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            36,
            6,
            32,
            83,
            143,
            50
          ]
        },
        {
          "session_id": "dea5acee-0685-4a83-a4af-38a5797d45a5",
          "game_id": "sc25",
          "game_version": "635fd71a",
          "published_at": "2026-04-14T17:43:47.463186Z",
          "score": 0,
          "model": "claude-opus-4-6",
          "runner": "benchmark_agent",
          "config": "anthropic-opus-4-6-max-effort",
          "verified_publish": true,
          "levels_completed": 0,
          "actions": 180,
          "resets": 11,
          "state": "NOT_FINISHED",
          "completed": false,
          "number_of_levels": 6,
          "level_scores": [
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            180,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            36,
            6,
            32,
            83,
            143,
            50
          ]
        }
      ]
    },
    {
      "game": "sb26",
      "source": "https://arcprize.org/tasks/sb26",
      "runs": [
        {
          "session_id": "4d7046ac-20d9-414e-a9e5-21a15c081675",
          "game_id": "sb26",
          "game_version": "7fbdac44",
          "published_at": "2026-09-25T06:39:03.866835Z",
          "score": 0.27700831024930744,
          "model": "gpt-6-sol",
          "runner": "benchmark_agent",
          "config": "openai-gpt-6-sol-none",
          "verified_publish": true,
          "levels_completed": 1,
          "actions": 197,
          "resets": 0,
          "state": "NOT_FINISHED",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            9.972299168975068,
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            57,
            140,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            18,
            28,
            18,
            19,
            31,
            23,
            58,
            18
          ]
        },
        {
          "session_id": "fa38cf10-8258-452a-b7f1-1394cb37a4a9",
          "game_id": "sb26",
          "game_version": "7fbdac44",
          "published_at": "2026-09-25T06:31:28.086491Z",
          "score": 0,
          "model": "gpt-6-sol",
          "runner": "benchmark_agent",
          "config": "openai-gpt-6-sol-none-provider-adapter",
          "verified_publish": true,
          "levels_completed": 0,
          "actions": 90,
          "resets": 1,
          "state": "NOT_FINISHED",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            0,
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            90,
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            18,
            28,
            18,
            19,
            31,
            23,
            58,
            18
          ]
        },
        {
          "session_id": "fdcee0ea-f524-44a7-9c9e-2c217d9f4324",
          "game_id": "sb26",
          "game_version": "7fbdac44",
          "published_at": "2026-09-25T06:05:51.118438Z",
          "score": 0,
          "model": "gpt-6-sol",
          "runner": "benchmark_agent",
          "config": "openai-gpt-6-sol-low-provider-adapter",
          "verified_publish": true,
          "levels_completed": 0,
          "actions": 90,
          "resets": 2,
          "state": "NOT_FINISHED",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            0,
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            90,
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            18,
            28,
            18,
            19,
            31,
            23,
            58,
            18
          ]
        },
        {
          "session_id": "2bebaf18-9d5f-484e-8d5c-eb02d17ce4d4",
          "game_id": "sb26",
          "game_version": "7fbdac44",
          "published_at": "2026-09-25T05:52:20.869586Z",
          "score": 0,
          "model": "gpt-6-sol",
          "runner": "benchmark_agent",
          "config": "openai-gpt-6-sol-low",
          "verified_publish": true,
          "levels_completed": 0,
          "actions": 90,
          "resets": 2,
          "state": "NOT_FINISHED",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            0,
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            90,
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            18,
            28,
            18,
            19,
            31,
            23,
            58,
            18
          ]
        },
        {
          "session_id": "8e9a83ff-88df-4413-a440-0fbd2eb352c7",
          "game_id": "sb26",
          "game_version": "7fbdac44",
          "published_at": "2026-09-25T05:45:13.543048Z",
          "score": 2.7777777777777777,
          "model": "gpt-6-sol",
          "runner": "benchmark_agent",
          "config": "openai-gpt-6-sol-medium-provider-adapter",
          "verified_publish": true,
          "levels_completed": 1,
          "actions": 158,
          "resets": 2,
          "state": "NOT_FINISHED",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            100,
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            18,
            140,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            18,
            28,
            18,
            19,
            31,
            23,
            58,
            18
          ]
        },
        {
          "session_id": "da30fea7-ed12-470a-a395-8be2bb3c0285",
          "game_id": "sb26",
          "game_version": "7fbdac44",
          "published_at": "2026-09-25T05:29:07.814631Z",
          "score": 0.5353955978584177,
          "model": "gpt-6-sol",
          "runner": "benchmark_agent",
          "config": "openai-gpt-6-sol-max",
          "verified_publish": true,
          "levels_completed": 1,
          "actions": 181,
          "resets": 0,
          "state": "NOT_FINISHED",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            19.274241522903036,
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            41,
            140,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            18,
            28,
            18,
            19,
            31,
            23,
            58,
            18
          ]
        },
        {
          "session_id": "be70c963-c4ac-46bd-b9b0-67472e94d9a1",
          "game_id": "sb26",
          "game_version": "7fbdac44",
          "published_at": "2026-09-25T05:07:47.547375Z",
          "score": 0,
          "model": "gpt-6-sol",
          "runner": "benchmark_agent",
          "config": "openai-gpt-6-sol-medium",
          "verified_publish": true,
          "levels_completed": 0,
          "actions": 90,
          "resets": 0,
          "state": "NOT_FINISHED",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            0,
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            90,
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            18,
            28,
            18,
            19,
            31,
            23,
            58,
            18
          ]
        },
        {
          "session_id": "53d4440b-8a48-40a2-bf75-bc568fd0f7dd",
          "game_id": "sb26",
          "game_version": "7fbdac44",
          "published_at": "2026-09-25T04:40:06.405154Z",
          "score": 2.7777777777777777,
          "model": "gpt-6-sol",
          "runner": "benchmark_agent",
          "config": "openai-gpt-6-sol-high",
          "verified_publish": true,
          "levels_completed": 1,
          "actions": 157,
          "resets": 0,
          "state": "NOT_FINISHED",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            112.11072664359862,
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            17,
            140,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            18,
            28,
            18,
            19,
            31,
            23,
            58,
            18
          ]
        },
        {
          "session_id": "30ef4e2b-095e-415d-ad86-071baa66636d",
          "game_id": "sb26",
          "game_version": "7fbdac44",
          "published_at": "2026-09-25T04:39:20.466849Z",
          "score": 0,
          "model": "gpt-6-sol",
          "runner": "benchmark_agent",
          "config": "openai-gpt-6-sol-high-provider-adapter",
          "verified_publish": true,
          "levels_completed": 0,
          "actions": 90,
          "resets": 2,
          "state": "NOT_FINISHED",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            0,
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            90,
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            18,
            28,
            18,
            19,
            31,
            23,
            58,
            18
          ]
        },
        {
          "session_id": "59d3b2e3-c635-4879-ab3d-7683f34276d9",
          "game_id": "sb26",
          "game_version": "7fbdac44",
          "published_at": "2026-09-25T04:05:04.013275Z",
          "score": 0,
          "model": "gpt-6-sol",
          "runner": "benchmark_agent",
          "config": "openai-gpt-6-sol-xhigh",
          "verified_publish": true,
          "levels_completed": 0,
          "actions": 90,
          "resets": 0,
          "state": "NOT_FINISHED",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            0,
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            90,
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            18,
            28,
            18,
            19,
            31,
            23,
            58,
            18
          ]
        },
        {
          "session_id": "fb2937fc-9acf-4fd5-8918-01006a90ed9e",
          "game_id": "sb26",
          "game_version": "7fbdac44",
          "published_at": "2026-09-25T04:02:00.338177Z",
          "score": 0,
          "model": "gpt-6-sol",
          "runner": "benchmark_agent",
          "config": "openai-gpt-6-sol-xhigh-provider-adapter",
          "verified_publish": true,
          "levels_completed": 0,
          "actions": 90,
          "resets": 0,
          "state": "NOT_FINISHED",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            0,
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            90,
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            18,
            28,
            18,
            19,
            31,
            23,
            58,
            18
          ]
        },
        {
          "session_id": "4dae0d23-d04b-4bd6-8676-5f1e5309dc88",
          "game_id": "sb26",
          "game_version": "7fbdac44",
          "published_at": "2026-09-21T12:13:18.200469Z",
          "score": 4.291838134430727,
          "model": "gemini-3.8-flash",
          "runner": "benchmark_agent",
          "config": "google-gemini-3-8-flash-low-provider-adapter",
          "verified_publish": true,
          "levels_completed": 2,
          "actions": 167,
          "resets": 1,
          "state": "NOT_FINISHED",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            115,
            19.753086419753085,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            14,
            63,
            90,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            18,
            28,
            18,
            19,
            31,
            23,
            58,
            18
          ]
        },
        {
          "session_id": "0eae7568-54b8-40b8-aa13-93ca89656eb6",
          "game_id": "sb26",
          "game_version": "7fbdac44",
          "published_at": "2026-09-20T08:17:19.931478Z",
          "score": 100,
          "model": "gemini-3.8-flash",
          "runner": "benchmark_agent",
          "config": "google-gemini-3-8-flash-medium-provider-adapter",
          "verified_publish": true,
          "levels_completed": 8,
          "actions": 146,
          "resets": 0,
          "state": "WIN",
          "completed": true,
          "number_of_levels": 0,
          "level_scores": [
            115,
            81.58168574401664,
            115,
            115,
            115,
            115,
            115,
            112.11072664359862
          ],
          "level_actions": [
            14,
            31,
            15,
            15,
            18,
            19,
            17,
            17
          ],
          "level_baseline_actions": [
            18,
            28,
            18,
            19,
            31,
            23,
            58,
            18
          ]
        },
        {
          "session_id": "9adee855-18a7-4c4e-b163-1b26f9b79d2c",
          "game_id": "sb26",
          "game_version": "7fbdac44",
          "published_at": "2026-09-17T19:00:12.930058Z",
          "score": 100,
          "model": "gemini-3.8-flash",
          "runner": "benchmark_agent",
          "config": "google-gemini-3-8-flash-high-provider-adapter",
          "verified_publish": true,
          "levels_completed": 8,
          "actions": 126,
          "resets": 0,
          "state": "WIN",
          "completed": true,
          "number_of_levels": 0,
          "level_scores": [
            115,
            115,
            115,
            115,
            115,
            115,
            115,
            112.11072664359862
          ],
          "level_actions": [
            10,
            15,
            15,
            16,
            17,
            19,
            17,
            17
          ],
          "level_baseline_actions": [
            18,
            28,
            18,
            19,
            31,
            23,
            58,
            18
          ]
        },
        {
          "session_id": "3e9e14f9-0dbf-4149-8cf8-f13c3026be39",
          "game_id": "sb26",
          "game_version": "7fbdac44",
          "published_at": "2026-09-16T19:25:48.871401Z",
          "score": 0,
          "model": "gpt-6-luna",
          "runner": "benchmark_agent",
          "config": "openai-gpt-6-luna-none-provider-adapter",
          "verified_publish": true,
          "levels_completed": 0,
          "actions": 90,
          "resets": 1,
          "state": "NOT_FINISHED",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            0,
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            90,
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            18,
            28,
            18,
            19,
            31,
            23,
            58,
            18
          ]
        },
        {
          "session_id": "365cf5df-5954-46ad-9ac7-934621544fab",
          "game_id": "sb26",
          "game_version": "7fbdac44",
          "published_at": "2026-09-16T19:24:00.363364Z",
          "score": 0,
          "model": "gpt-6-luna",
          "runner": "benchmark_agent",
          "config": "openai-gpt-6-luna-low",
          "verified_publish": true,
          "levels_completed": 0,
          "actions": 90,
          "resets": 1,
          "state": "NOT_FINISHED",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            0,
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            90,
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            18,
            28,
            18,
            19,
            31,
            23,
            58,
            18
          ]
        },
        {
          "session_id": "baf1fbf7-a3cb-4c2e-9f57-82ddc09677d2",
          "game_id": "sb26",
          "game_version": "7fbdac44",
          "published_at": "2026-09-16T19:23:51.08499Z",
          "score": 0,
          "model": "gpt-6-luna",
          "runner": "benchmark_agent",
          "config": "openai-gpt-6-luna-low-provider-adapter",
          "verified_publish": true,
          "levels_completed": 0,
          "actions": 90,
          "resets": 0,
          "state": "NOT_FINISHED",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            0,
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            90,
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            18,
            28,
            18,
            19,
            31,
            23,
            58,
            18
          ]
        },
        {
          "session_id": "8bebce5a-4d9c-45b8-91f1-2a42740d74b1",
          "game_id": "sb26",
          "game_version": "7fbdac44",
          "published_at": "2026-09-16T19:22:17.169426Z",
          "score": 0,
          "model": "gpt-6-luna",
          "runner": "benchmark_agent",
          "config": "openai-gpt-6-luna-none",
          "verified_publish": true,
          "levels_completed": 0,
          "actions": 90,
          "resets": 1,
          "state": "NOT_FINISHED",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            0,
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            90,
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            18,
            28,
            18,
            19,
            31,
            23,
            58,
            18
          ]
        },
        {
          "session_id": "ea335387-d4d5-4f5d-b2ee-7ca5123d0401",
          "game_id": "sb26",
          "game_version": "7fbdac44",
          "published_at": "2026-09-16T19:21:55.116136Z",
          "score": 2.7777777777777777,
          "model": "gpt-6-1-sol",
          "runner": "benchmark_agent",
          "config": "openai-gpt-6-1-sol-low",
          "verified_publish": true,
          "levels_completed": 1,
          "actions": 149,
          "resets": 0,
          "state": "NOT_FINISHED",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            115,
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            9,
            140,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            18,
            28,
            18,
            19,
            31,
            23,
            58,
            18
          ]
        },
        {
          "session_id": "716ff325-9771-4bf0-bfef-28abc7b259d7",
          "game_id": "sb26",
          "game_version": "7fbdac44",
          "published_at": "2026-09-16T15:08:32.531296Z",
          "score": 0,
          "model": "gpt-6-luna",
          "runner": "benchmark_agent",
          "config": "openai-gpt-6-luna-xhigh",
          "verified_publish": true,
          "levels_completed": 0,
          "actions": 90,
          "resets": 0,
          "state": "NOT_FINISHED",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            0,
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            90,
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            18,
            28,
            18,
            19,
            31,
            23,
            58,
            18
          ]
        },
        {
          "session_id": "664f3466-dde4-45a9-b10d-4221da9cc7da",
          "game_id": "sb26",
          "game_version": "7fbdac44",
          "published_at": "2026-09-16T09:01:37.84951Z",
          "score": 0,
          "model": "gpt-6-luna",
          "runner": "benchmark_agent",
          "config": "openai-gpt-6-luna-medium-provider-adapter",
          "verified_publish": true,
          "levels_completed": 0,
          "actions": 90,
          "resets": 0,
          "state": "NOT_FINISHED",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            0,
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            90,
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            18,
            28,
            18,
            19,
            31,
            23,
            58,
            18
          ]
        },
        {
          "session_id": "45a8ca51-e12a-4c0a-86bc-1d01f76f3446",
          "game_id": "sb26",
          "game_version": "7fbdac44",
          "published_at": "2026-09-16T08:52:28.354327Z",
          "score": 0,
          "model": "gpt-6-luna",
          "runner": "benchmark_agent",
          "config": "openai-gpt-6-luna-medium",
          "verified_publish": true,
          "levels_completed": 0,
          "actions": 90,
          "resets": 0,
          "state": "NOT_FINISHED",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            0,
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            90,
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            18,
            28,
            18,
            19,
            31,
            23,
            58,
            18
          ]
        },
        {
          "session_id": "33c03f48-98d1-43da-ac1e-36f566eddc9b",
          "game_id": "sb26",
          "game_version": "7fbdac44",
          "published_at": "2026-09-16T08:25:42.858249Z",
          "score": 0,
          "model": "gpt-6-luna",
          "runner": "benchmark_agent",
          "config": "openai-gpt-6-luna-high-provider-adapter",
          "verified_publish": true,
          "levels_completed": 0,
          "actions": 90,
          "resets": 1,
          "state": "NOT_FINISHED",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            0,
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            90,
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            18,
            28,
            18,
            19,
            31,
            23,
            58,
            18
          ]
        },
        {
          "session_id": "112aaa37-7300-4c2b-8af5-8b11d4f3d66b",
          "game_id": "sb26",
          "game_version": "7fbdac44",
          "published_at": "2026-09-16T08:21:04.053599Z",
          "score": 0,
          "model": "gpt-6-luna",
          "runner": "benchmark_agent",
          "config": "openai-gpt-6-luna-high",
          "verified_publish": true,
          "levels_completed": 0,
          "actions": 90,
          "resets": 0,
          "state": "NOT_FINISHED",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            0,
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            90,
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            18,
            28,
            18,
            19,
            31,
            23,
            58,
            18
          ]
        },
        {
          "session_id": "0b6eba92-d687-4f3f-ab9a-b25bdfba9e9e",
          "game_id": "sb26",
          "game_version": "7fbdac44",
          "published_at": "2026-09-16T07:30:18.326631Z",
          "score": 0,
          "model": "gpt-6-luna",
          "runner": "benchmark_agent",
          "config": "openai-gpt-6-luna-xhigh-provider-adapter",
          "verified_publish": true,
          "levels_completed": 0,
          "actions": 90,
          "resets": 0,
          "state": "NOT_FINISHED",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            0,
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            90,
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            18,
            28,
            18,
            19,
            31,
            23,
            58,
            18
          ]
        },
        {
          "session_id": "8039c117-6a72-4b49-8524-7c612695beef",
          "game_id": "sb26",
          "game_version": "7fbdac44",
          "published_at": "2026-09-16T06:46:10.99253Z",
          "score": 0,
          "model": "gpt-6-luna",
          "runner": "benchmark_agent",
          "config": "openai-gpt-6-luna-max-provider-adapter",
          "verified_publish": true,
          "levels_completed": 0,
          "actions": 90,
          "resets": 1,
          "state": "NOT_FINISHED",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            0,
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            90,
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            18,
            28,
            18,
            19,
            31,
            23,
            58,
            18
          ]
        },
        {
          "session_id": "cbbecde3-c1c2-4dc7-850c-e0eab397400e",
          "game_id": "sb26",
          "game_version": "7fbdac44",
          "published_at": "2026-09-16T06:20:22.009727Z",
          "score": 0,
          "model": "gpt-6-luna",
          "runner": "benchmark_agent",
          "config": "openai-gpt-6-luna-max",
          "verified_publish": true,
          "levels_completed": 0,
          "actions": 90,
          "resets": 0,
          "state": "NOT_FINISHED",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            0,
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            90,
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            18,
            28,
            18,
            19,
            31,
            23,
            58,
            18
          ]
        },
        {
          "session_id": "ca8fbd19-a7fe-440a-98ec-c688a1439f34",
          "game_id": "sb26",
          "game_version": "7fbdac44",
          "published_at": "2026-09-16T05:25:50.49475Z",
          "score": 58.842907479929686,
          "model": "gpt-6-1-sol",
          "runner": "benchmark_agent",
          "config": "openai-gpt-6-1-sol-medium",
          "verified_publish": true,
          "levels_completed": 7,
          "actions": 315,
          "resets": 0,
          "state": "NOT_FINISHED",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            115,
            22.52226371732261,
            89.75069252077562,
            115,
            38.440000000000005,
            38.641344046749445,
            115,
            0
          ],
          "level_actions": [
            9,
            59,
            19,
            17,
            50,
            37,
            34,
            90
          ],
          "level_baseline_actions": [
            18,
            28,
            18,
            19,
            31,
            23,
            58,
            18
          ]
        },
        {
          "session_id": "817f4b81-89f8-46e2-9c0c-f205dd3b5c45",
          "game_id": "sb26",
          "game_version": "7fbdac44",
          "published_at": "2026-09-16T04:22:42.431293Z",
          "score": 100,
          "model": "gpt-6-1-sol",
          "runner": "benchmark_agent",
          "config": "openai-gpt-6-1-sol-high-provider-adapter",
          "verified_publish": true,
          "levels_completed": 8,
          "actions": 125,
          "resets": 0,
          "state": "WIN",
          "completed": true,
          "number_of_levels": 0,
          "level_scores": [
            115,
            115,
            115,
            115,
            115,
            115,
            115,
            112.11072664359862
          ],
          "level_actions": [
            10,
            15,
            15,
            15,
            17,
            19,
            17,
            17
          ],
          "level_baseline_actions": [
            18,
            28,
            18,
            19,
            31,
            23,
            58,
            18
          ]
        },
        {
          "session_id": "1bbef032-a12a-4ba7-9100-f143730582ec",
          "game_id": "sb26",
          "game_version": "7fbdac44",
          "published_at": "2026-09-16T03:46:12.526217Z",
          "score": 100,
          "model": "gpt-6-1-sol",
          "runner": "benchmark_agent",
          "config": "openai-gpt-6-1-sol-high",
          "verified_publish": true,
          "levels_completed": 8,
          "actions": 136,
          "resets": 0,
          "state": "WIN",
          "completed": true,
          "number_of_levels": 0,
          "level_scores": [
            115,
            115,
            115,
            115,
            115,
            91.84027777777779,
            115,
            89.75069252077562
          ],
          "level_actions": [
            10,
            16,
            15,
            15,
            19,
            24,
            18,
            19
          ],
          "level_baseline_actions": [
            18,
            28,
            18,
            19,
            31,
            23,
            58,
            18
          ]
        },
        {
          "session_id": "533b0be7-85a7-49d0-98ba-f6e45fe135a2",
          "game_id": "sb26",
          "game_version": "7fbdac44",
          "published_at": "2026-09-16T03:36:13.199162Z",
          "score": 100,
          "model": "gpt-6-1-sol",
          "runner": "benchmark_agent",
          "config": "openai-gpt-6-1-sol-max",
          "verified_publish": true,
          "levels_completed": 8,
          "actions": 124,
          "resets": 0,
          "state": "WIN",
          "completed": true,
          "number_of_levels": 0,
          "level_scores": [
            115,
            115,
            115,
            115,
            115,
            115,
            115,
            112.11072664359862
          ],
          "level_actions": [
            9,
            15,
            15,
            15,
            17,
            19,
            17,
            17
          ],
          "level_baseline_actions": [
            18,
            28,
            18,
            19,
            31,
            23,
            58,
            18
          ]
        },
        {
          "session_id": "204adc51-dcea-442f-8c17-027a131e0080",
          "game_id": "sb26",
          "game_version": "7fbdac44",
          "published_at": "2026-09-16T03:21:31.339506Z",
          "score": 100,
          "model": "gpt-6-1-sol",
          "runner": "benchmark_agent",
          "config": "openai-gpt-6-1-sol-max-provider-adapter",
          "verified_publish": true,
          "levels_completed": 8,
          "actions": 124,
          "resets": 0,
          "state": "WIN",
          "completed": true,
          "number_of_levels": 0,
          "level_scores": [
            115,
            115,
            115,
            115,
            115,
            115,
            115,
            112.11072664359862
          ],
          "level_actions": [
            9,
            15,
            15,
            15,
            17,
            19,
            17,
            17
          ],
          "level_baseline_actions": [
            18,
            28,
            18,
            19,
            31,
            23,
            58,
            18
          ]
        },
        {
          "session_id": "d39e626b-1514-488b-85ac-17b10fa4d68c",
          "game_id": "sb26",
          "game_version": "7fbdac44",
          "published_at": "2026-09-16T03:21:13.598261Z",
          "score": 100,
          "model": "gpt-6-1-sol",
          "runner": "benchmark_agent",
          "config": "openai-gpt-6-1-sol-xhigh",
          "verified_publish": true,
          "levels_completed": 8,
          "actions": 124,
          "resets": 0,
          "state": "WIN",
          "completed": true,
          "number_of_levels": 0,
          "level_scores": [
            115,
            115,
            115,
            115,
            115,
            115,
            115,
            112.11072664359862
          ],
          "level_actions": [
            9,
            15,
            15,
            15,
            17,
            19,
            17,
            17
          ],
          "level_baseline_actions": [
            18,
            28,
            18,
            19,
            31,
            23,
            58,
            18
          ]
        },
        {
          "session_id": "efc00249-4ec7-4b95-9c44-22295283604e",
          "game_id": "sb26",
          "game_version": "7fbdac44",
          "published_at": "2026-09-16T02:47:21.342522Z",
          "score": 100,
          "model": "gpt-6-1-sol",
          "runner": "benchmark_agent",
          "config": "openai-gpt-6-1-sol-xhigh-provider-adapter",
          "verified_publish": true,
          "levels_completed": 8,
          "actions": 124,
          "resets": 0,
          "state": "WIN",
          "completed": true,
          "number_of_levels": 0,
          "level_scores": [
            115,
            115,
            115,
            115,
            115,
            115,
            115,
            112.11072664359862
          ],
          "level_actions": [
            9,
            15,
            15,
            15,
            17,
            19,
            17,
            17
          ],
          "level_baseline_actions": [
            18,
            28,
            18,
            19,
            31,
            23,
            58,
            18
          ]
        },
        {
          "session_id": "0b521ec6-c838-4fe8-86d2-2b38f97b2afa",
          "game_id": "sb26",
          "game_version": "7fbdac44",
          "published_at": "2026-09-12T11:49:48.921448Z",
          "score": 2.7777777777777777,
          "model": "gemini-3.8-flash",
          "runner": "benchmark_agent",
          "config": "google-gemini-3-8-flash-low",
          "verified_publish": true,
          "levels_completed": 1,
          "actions": 149,
          "resets": 1,
          "state": "NOT_FINISHED",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            115,
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            9,
            140,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            18,
            28,
            18,
            19,
            31,
            23,
            58,
            18
          ]
        },
        {
          "session_id": "fbf1754b-ef3a-497d-9967-3e527726798d",
          "game_id": "sb26",
          "game_version": "7fbdac44",
          "published_at": "2026-09-12T05:53:39.047324Z",
          "score": 2.7777777777777777,
          "model": "gemini-3.8-flash",
          "runner": "benchmark_agent",
          "config": "google-gemini-3-8-flash-medium",
          "verified_publish": true,
          "levels_completed": 1,
          "actions": 149,
          "resets": 3,
          "state": "NOT_FINISHED",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            115,
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            9,
            140,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            18,
            28,
            18,
            19,
            31,
            23,
            58,
            18
          ]
        },
        {
          "session_id": "ccfce96b-7c5c-44d8-9f57-2e073634ca9b",
          "game_id": "sb26",
          "game_version": "7fbdac44",
          "published_at": "2026-09-09T17:27:32.209027Z",
          "score": 2.7777777777777777,
          "model": "gemini-3.8-flash",
          "runner": "benchmark_agent",
          "config": "google-gemini-3-8-flash-high",
          "verified_publish": true,
          "levels_completed": 1,
          "actions": 156,
          "resets": 0,
          "state": "NOT_FINISHED",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            115,
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            16,
            140,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            18,
            28,
            18,
            19,
            31,
            23,
            58,
            18
          ]
        },
        {
          "session_id": "5385832e-3742-4b44-a163-7c9b41a7b415",
          "game_id": "sb26",
          "game_version": "7fbdac44",
          "published_at": "2026-08-11T07:08:22.049074Z",
          "score": 2.7777777777777777,
          "model": "grok-4.6",
          "runner": "benchmark_agent",
          "config": "xai-grok-4-6-xhigh",
          "verified_publish": true,
          "levels_completed": 1,
          "actions": 158,
          "resets": 3,
          "state": "NOT_FINISHED",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            100,
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            18,
            140,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            18,
            28,
            18,
            19,
            31,
            23,
            58,
            18
          ]
        },
        {
          "session_id": "f19392e3-15ba-4ff2-867c-d00ce797e360",
          "game_id": "sb26",
          "game_version": "7fbdac44",
          "published_at": "2026-07-21T01:00:17.372576Z",
          "score": 77.77777777777779,
          "model": "claude-opus-5",
          "runner": "benchmark_agent",
          "config": "anthropic-claude-opus-5-high",
          "verified_publish": true,
          "levels_completed": 7,
          "actions": 200,
          "resets": 0,
          "state": "NOT_FINISHED",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            115,
            115,
            115,
            115,
            115,
            115,
            115,
            0
          ],
          "level_actions": [
            9,
            15,
            16,
            15,
            19,
            19,
            17,
            90
          ],
          "level_baseline_actions": [
            18,
            28,
            18,
            19,
            31,
            23,
            58,
            18
          ]
        },
        {
          "session_id": "827f7a86-84fb-44eb-a30a-bfa5e2f22238",
          "game_id": "sb26",
          "game_version": "7fbdac44",
          "published_at": "2026-06-24T10:45:10.836396Z",
          "score": 0,
          "model": "gpt-5-6-luna",
          "runner": "benchmark_agent",
          "config": "openai-gpt-5-6-luna-low",
          "verified_publish": true,
          "levels_completed": 0,
          "actions": 90,
          "resets": 0,
          "state": "NOT_FINISHED",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            0,
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            90,
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            18,
            28,
            18,
            19,
            31,
            23,
            58,
            18
          ]
        },
        {
          "session_id": "bb161a3e-65e9-42fc-8d0c-d64817cf7076",
          "game_id": "sb26",
          "game_version": "7fbdac44",
          "published_at": "2026-06-24T10:22:17.264907Z",
          "score": 0,
          "model": "gpt-5-6-luna",
          "runner": "benchmark_agent",
          "config": "openai-gpt-5-6-luna-medium",
          "verified_publish": true,
          "levels_completed": 0,
          "actions": 90,
          "resets": 1,
          "state": "NOT_FINISHED",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            0,
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            90,
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            18,
            28,
            18,
            19,
            31,
            23,
            58,
            18
          ]
        },
        {
          "session_id": "21740bc5-811e-4416-a67f-22907887596d",
          "game_id": "sb26",
          "game_version": "7fbdac44",
          "published_at": "2026-06-24T03:09:23.875108Z",
          "score": 0,
          "model": "gpt-5-6-luna",
          "runner": "benchmark_agent",
          "config": "openai-gpt-5-6-luna-high",
          "verified_publish": true,
          "levels_completed": 0,
          "actions": 90,
          "resets": 1,
          "state": "NOT_FINISHED",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            0,
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            90,
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            18,
            28,
            18,
            19,
            31,
            23,
            58,
            18
          ]
        },
        {
          "session_id": "226f1b70-f6cb-4c2a-8d92-d0a43c5eed4d",
          "game_id": "sb26",
          "game_version": "7fbdac44",
          "published_at": "2026-06-23T23:42:53.935221Z",
          "score": 0,
          "model": "gpt-5-6-luna",
          "runner": "benchmark_agent",
          "config": "openai-gpt-5-6-luna-xhigh",
          "verified_publish": true,
          "levels_completed": 0,
          "actions": 90,
          "resets": 5,
          "state": "NOT_FINISHED",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            0,
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            90,
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            18,
            28,
            18,
            19,
            31,
            23,
            58,
            18
          ]
        },
        {
          "session_id": "0f9a03f7-a3e8-4744-a9d8-b45fb18a875d",
          "game_id": "sb26",
          "game_version": "7fbdac44",
          "published_at": "2026-06-23T00:44:01.801Z",
          "score": 0,
          "model": "gpt-5-6-luna",
          "runner": "benchmark_agent",
          "config": "openai-gpt-5-6-luna-max",
          "verified_publish": true,
          "levels_completed": 0,
          "actions": 90,
          "resets": 6,
          "state": "NOT_FINISHED",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            0,
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            90,
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            18,
            28,
            18,
            19,
            31,
            23,
            58,
            18
          ]
        },
        {
          "session_id": "2d417444-871b-4c78-a756-8853d79ae4db",
          "game_id": "sb26",
          "game_version": "7fbdac44",
          "published_at": "2026-06-20T11:54:09.775245Z",
          "score": 0,
          "model": "gpt-5-6-terra",
          "runner": "benchmark_agent",
          "config": "openai-gpt-5-6-terra-low",
          "verified_publish": true,
          "levels_completed": 0,
          "actions": 90,
          "resets": 0,
          "state": "NOT_FINISHED",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            0,
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            90,
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            18,
            28,
            18,
            19,
            31,
            23,
            58,
            18
          ]
        },
        {
          "session_id": "5835ac81-c421-42ad-952d-7858237431a2",
          "game_id": "sb26",
          "game_version": "7fbdac44",
          "published_at": "2026-06-20T09:08:01.302319Z",
          "score": 0,
          "model": "gpt-5-6-terra",
          "runner": "benchmark_agent",
          "config": "openai-gpt-5-6-terra-medium",
          "verified_publish": true,
          "levels_completed": 0,
          "actions": 90,
          "resets": 0,
          "state": "NOT_FINISHED",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            0,
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            90,
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            18,
            28,
            18,
            19,
            31,
            23,
            58,
            18
          ]
        },
        {
          "session_id": "25a4457d-e6f9-4d3d-871f-cc3cd18e6203",
          "game_id": "sb26",
          "game_version": "7fbdac44",
          "published_at": "2026-06-20T08:39:40.744694Z",
          "score": 0.3748438150770512,
          "model": "gpt-5-6-terra",
          "runner": "benchmark_agent",
          "config": "openai-gpt-5-6-terra-high",
          "verified_publish": true,
          "levels_completed": 1,
          "actions": 189,
          "resets": 1,
          "state": "NOT_FINISHED",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            13.494377342773845,
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            49,
            140,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            18,
            28,
            18,
            19,
            31,
            23,
            58,
            18
          ]
        },
        {
          "session_id": "33af9d91-3a96-4dcf-a07b-bdd9e4b38f1a",
          "game_id": "sb26",
          "game_version": "7fbdac44",
          "published_at": "2026-06-20T01:56:13.717008Z",
          "score": 0.3748438150770512,
          "model": "gpt-5-6-terra",
          "runner": "benchmark_agent",
          "config": "openai-gpt-5-6-terra-xhigh",
          "verified_publish": true,
          "levels_completed": 1,
          "actions": 189,
          "resets": 4,
          "state": "NOT_FINISHED",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            13.494377342773845,
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            49,
            140,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            18,
            28,
            18,
            19,
            31,
            23,
            58,
            18
          ]
        },
        {
          "session_id": "2076f400-aa16-4f99-8e50-f1b8ae4fa723",
          "game_id": "sb26",
          "game_version": "7fbdac44",
          "published_at": "2026-06-18T20:50:39.958264Z",
          "score": 0,
          "model": "gpt-5-6-terra",
          "runner": "benchmark_agent",
          "config": "openai-gpt-5-6-terra-max",
          "verified_publish": true,
          "levels_completed": 0,
          "actions": 90,
          "resets": 0,
          "state": "NOT_FINISHED",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            0,
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            90,
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            18,
            28,
            18,
            19,
            31,
            23,
            58,
            18
          ]
        },
        {
          "session_id": "5e554fc3-4bde-43ac-bb8e-738969e7e67a",
          "game_id": "sb26",
          "game_version": "7fbdac44",
          "published_at": "2026-06-13T02:50:35.550261Z",
          "score": 0.21301775147928997,
          "model": "gpt-5-6-sol",
          "runner": "benchmark_agent",
          "config": "openai-gpt-5-6-sol-max",
          "verified_publish": true,
          "levels_completed": 1,
          "actions": 205,
          "resets": 3,
          "state": "NOT_FINISHED",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            7.668639053254439,
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            65,
            140,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            18,
            28,
            18,
            19,
            31,
            23,
            58,
            18
          ]
        },
        {
          "session_id": "a958a3bf-d26b-424c-a6bc-4d1715029611",
          "game_id": "sb26",
          "game_version": "7fbdac44",
          "published_at": "2026-06-11T13:37:58.44135Z",
          "score": 0,
          "model": "gpt-5-6-sol",
          "runner": "benchmark_agent",
          "config": "openai-gpt-5-6-sol-xhigh",
          "verified_publish": true,
          "levels_completed": 0,
          "actions": 90,
          "resets": 1,
          "state": "NOT_FINISHED",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            0,
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            90,
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            18,
            28,
            18,
            19,
            31,
            23,
            58,
            18
          ]
        },
        {
          "session_id": "39809b8a-4db4-4550-8199-13ca0f272f03",
          "game_id": "sb26",
          "game_version": "7fbdac44",
          "published_at": "2026-06-10T10:58:32.823955Z",
          "score": 0.4253308128544424,
          "model": "gpt-5-6-sol",
          "runner": "benchmark_agent",
          "config": "openai-gpt-5-6-sol-high",
          "verified_publish": true,
          "levels_completed": 1,
          "actions": 186,
          "resets": 3,
          "state": "NOT_FINISHED",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            15.311909262759926,
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            46,
            140,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            18,
            28,
            18,
            19,
            31,
            23,
            58,
            18
          ]
        },
        {
          "session_id": "a054edfe-4590-45f7-a4c6-b9218ad48e2c",
          "game_id": "sb26",
          "game_version": "7fbdac44",
          "published_at": "2026-06-10T07:46:31.071621Z",
          "score": 0,
          "model": "gpt-5-6-sol",
          "runner": "benchmark_agent",
          "config": "openai-gpt-5-6-sol-medium",
          "verified_publish": true,
          "levels_completed": 0,
          "actions": 90,
          "resets": 0,
          "state": "NOT_FINISHED",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            0,
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            90,
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            18,
            28,
            18,
            19,
            31,
            23,
            58,
            18
          ]
        },
        {
          "session_id": "9e4368ee-bf70-4bd2-852b-c0f48c864317",
          "game_id": "sb26",
          "game_version": "7fbdac44",
          "published_at": "2026-06-09T21:23:03.9524Z",
          "score": 0,
          "model": "gpt-5-6-sol",
          "runner": "benchmark_agent",
          "config": "openai-gpt-5-6-sol-low",
          "verified_publish": true,
          "levels_completed": 0,
          "actions": 90,
          "resets": 3,
          "state": "NOT_FINISHED",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            0,
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            90,
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            18,
            28,
            18,
            19,
            31,
            23,
            58,
            18
          ]
        },
        {
          "session_id": "e81e1eed-99d1-4199-b19e-b0ee10442f26",
          "game_id": "sb26",
          "game_version": "7fbdac44",
          "published_at": "2026-05-29T17:12:55.647974Z",
          "score": 2.7777777777777777,
          "model": "claude-opus-4-8",
          "runner": "benchmark_agent",
          "config": "anthropic-opus-4-8-high",
          "verified_publish": true,
          "levels_completed": 1,
          "actions": 149,
          "resets": 4,
          "state": "NOT_FINISHED",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            115,
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            9,
            140,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            18,
            28,
            18,
            19,
            31,
            23,
            58,
            18
          ]
        },
        {
          "session_id": "4505fe65-5e17-4532-b71c-bd7c0dff58eb",
          "game_id": "sb26",
          "game_version": "7fbdac44",
          "published_at": "2026-05-01T05:07:04.376394Z",
          "score": 0,
          "model": "claude-opus-4-7",
          "runner": "benchmark_agent",
          "config": "anthropic-opus-4-7-high",
          "verified_publish": true,
          "levels_completed": 0,
          "actions": 90,
          "resets": 0,
          "state": "NOT_FINISHED",
          "completed": false,
          "number_of_levels": 8,
          "level_scores": [
            0,
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            90,
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            18,
            28,
            18,
            19,
            31,
            23,
            58,
            18
          ]
        },
        {
          "session_id": "31f9ed59-9318-4147-b9bf-6075519b8c51",
          "game_id": "sb26",
          "game_version": "7fbdac44",
          "published_at": "2026-04-30T23:20:23.116674Z",
          "score": 0,
          "model": "gpt-5.5-2026-04-23",
          "runner": "benchmark_agent",
          "config": "openai-gpt-5-5-2026-04-23-high",
          "verified_publish": true,
          "levels_completed": 0,
          "actions": 90,
          "resets": 1,
          "state": "NOT_FINISHED",
          "completed": false,
          "number_of_levels": 8,
          "level_scores": [
            0,
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            90,
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            18,
            28,
            18,
            19,
            31,
            23,
            58,
            18
          ]
        },
        {
          "session_id": "9b0fef90-f465-4f77-a899-ddbe0fa253d6",
          "game_id": "sb26",
          "game_version": "7fbdac44",
          "published_at": "2026-04-14T17:43:58.067879Z",
          "score": 0,
          "model": "grok-4.20-beta-0309-reasoning",
          "runner": "benchmark_agent",
          "config": "xai-grok-4-20-beta-0309-reasoning",
          "verified_publish": true,
          "levels_completed": 0,
          "actions": 74,
          "resets": 3,
          "state": "NOT_FINISHED",
          "completed": false,
          "number_of_levels": 8,
          "level_scores": [
            0,
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            74,
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            18,
            28,
            18,
            19,
            31,
            23,
            58,
            18
          ]
        },
        {
          "session_id": "e04cded9-3391-410f-9e52-67e11ccf62bf",
          "game_id": "sb26",
          "game_version": "7fbdac44",
          "published_at": "2026-04-14T17:43:57.522415Z",
          "score": 0,
          "model": "gpt-5.4-2026-03-05",
          "runner": "benchmark_agent",
          "config": "openai-gpt-5-4-2026-03-05-high",
          "verified_publish": true,
          "levels_completed": 0,
          "actions": 90,
          "resets": 2,
          "state": "NOT_FINISHED",
          "completed": false,
          "number_of_levels": 8,
          "level_scores": [
            0,
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            90,
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            18,
            28,
            18,
            19,
            31,
            23,
            58,
            18
          ]
        },
        {
          "session_id": "c30f2f34-d1eb-485d-a2a4-d3397b865387",
          "game_id": "sb26",
          "game_version": "7fbdac44",
          "published_at": "2026-04-14T17:43:56.14138Z",
          "score": 0.20661157024793386,
          "model": "gemini-3.1-pro-preview",
          "runner": "benchmark_agent",
          "config": "google-gemini-3-1-pro-preview",
          "verified_publish": true,
          "levels_completed": 1,
          "actions": 206,
          "resets": 5,
          "state": "NOT_FINISHED",
          "completed": false,
          "number_of_levels": 8,
          "level_scores": [
            7.438016528925619,
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            66,
            140,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            18,
            28,
            18,
            19,
            31,
            23,
            58,
            18
          ]
        },
        {
          "session_id": "b4c34b89-47f7-4dda-92d3-4973e4e3afb5",
          "game_id": "sb26",
          "game_version": "7fbdac44",
          "published_at": "2026-04-14T17:43:55.329661Z",
          "score": 0.19463667820069205,
          "model": "claude-opus-4-6",
          "runner": "benchmark_agent",
          "config": "anthropic-opus-4-6-max-effort",
          "verified_publish": true,
          "levels_completed": 1,
          "actions": 208,
          "resets": 2,
          "state": "NOT_FINISHED",
          "completed": false,
          "number_of_levels": 8,
          "level_scores": [
            7.006920415224914,
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            68,
            140,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            18,
            28,
            18,
            19,
            31,
            23,
            58,
            18
          ]
        }
      ]
    },
    {
      "game": "vc33",
      "source": "https://arcprize.org/tasks/vc33",
      "runs": [
        {
          "session_id": "d53fbd1b-b4ac-480d-8c31-0aa69553987a",
          "game_id": "vc33",
          "game_version": "5430563c",
          "published_at": "2026-09-25T15:26:30.492247Z",
          "score": 5.931521782188972,
          "model": "gpt-6-sol",
          "runner": "benchmark_agent",
          "config": "openai-gpt-6-sol-max",
          "verified_publish": true,
          "levels_completed": 3,
          "actions": 374,
          "resets": 5,
          "state": "GAME_OVER",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            7.2485207100591715,
            6.25,
            48.77802973041068,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            26,
            72,
            63,
            213,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            7,
            18,
            44,
            61,
            131,
            34,
            152
          ]
        },
        {
          "session_id": "c1ff6d06-b78c-4395-ac8a-c79988564f44",
          "game_id": "vc33",
          "game_version": "5430563c",
          "published_at": "2026-09-25T06:28:48.825852Z",
          "score": 0,
          "model": "gpt-6-sol",
          "runner": "benchmark_agent",
          "config": "openai-gpt-6-sol-none-provider-adapter",
          "verified_publish": true,
          "levels_completed": 0,
          "actions": 35,
          "resets": 0,
          "state": "NOT_FINISHED",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            35,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            7,
            18,
            44,
            61,
            131,
            34,
            152
          ]
        },
        {
          "session_id": "3403e221-74fd-45a6-bb5e-809584fa34d8",
          "game_id": "vc33",
          "game_version": "5430563c",
          "published_at": "2026-09-25T06:20:45.432441Z",
          "score": 0,
          "model": "gpt-6-sol",
          "runner": "benchmark_agent",
          "config": "openai-gpt-6-sol-none",
          "verified_publish": true,
          "levels_completed": 0,
          "actions": 35,
          "resets": 0,
          "state": "NOT_FINISHED",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            35,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            7,
            18,
            44,
            61,
            131,
            34,
            152
          ]
        },
        {
          "session_id": "5093d57e-bdcf-436e-8320-949c5cedc81c",
          "game_id": "vc33",
          "game_version": "5430563c",
          "published_at": "2026-09-25T06:03:58.379687Z",
          "score": 0,
          "model": "gpt-6-sol",
          "runner": "benchmark_agent",
          "config": "openai-gpt-6-sol-low-provider-adapter",
          "verified_publish": true,
          "levels_completed": 0,
          "actions": 35,
          "resets": 0,
          "state": "NOT_FINISHED",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            35,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            7,
            18,
            44,
            61,
            131,
            34,
            152
          ]
        },
        {
          "session_id": "dd4bb1b4-5dd5-4c97-9c48-0d53adc3922d",
          "game_id": "vc33",
          "game_version": "5430563c",
          "published_at": "2026-09-25T05:58:30.433692Z",
          "score": 0.1708984375,
          "model": "gpt-6-sol",
          "runner": "benchmark_agent",
          "config": "openai-gpt-6-sol-low",
          "verified_publish": true,
          "levels_completed": 1,
          "actions": 122,
          "resets": 1,
          "state": "GAME_OVER",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            4.78515625,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            32,
            90,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            7,
            18,
            44,
            61,
            131,
            34,
            152
          ]
        },
        {
          "session_id": "e8b05577-5136-4246-96d3-9e510edeacc1",
          "game_id": "vc33",
          "game_version": "5430563c",
          "published_at": "2026-09-25T05:38:22.552713Z",
          "score": 0,
          "model": "gpt-6-sol",
          "runner": "benchmark_agent",
          "config": "openai-gpt-6-sol-medium-provider-adapter",
          "verified_publish": true,
          "levels_completed": 0,
          "actions": 35,
          "resets": 0,
          "state": "NOT_FINISHED",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            35,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            7,
            18,
            44,
            61,
            131,
            34,
            152
          ]
        },
        {
          "session_id": "6cfe953e-f9ee-42e5-ab66-dcc0a17d1052",
          "game_id": "vc33",
          "game_version": "5430563c",
          "published_at": "2026-09-25T05:34:45.408464Z",
          "score": 10.76818783068783,
          "model": "gpt-6-sol",
          "runner": "benchmark_agent",
          "config": "openai-gpt-6-sol-xhigh-provider-adapter",
          "verified_publish": true,
          "levels_completed": 3,
          "actions": 433,
          "resets": 12,
          "state": "GAME_OVER",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            12.249999999999998,
            115,
            19.753086419753085,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            20,
            9,
            99,
            305,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            7,
            18,
            44,
            61,
            131,
            34,
            152
          ]
        },
        {
          "session_id": "bcda3431-1f9a-42c0-9b35-bb878a4a00be",
          "game_id": "vc33",
          "game_version": "5430563c",
          "published_at": "2026-09-25T05:18:32.739714Z",
          "score": 0,
          "model": "gpt-6-sol",
          "runner": "benchmark_agent",
          "config": "openai-gpt-6-sol-medium",
          "verified_publish": true,
          "levels_completed": 0,
          "actions": 35,
          "resets": 1,
          "state": "NOT_FINISHED",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            35,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            7,
            18,
            44,
            61,
            131,
            34,
            152
          ]
        },
        {
          "session_id": "a1808c9b-cfa3-42d6-a85d-c152bc11a6c0",
          "game_id": "vc33",
          "game_version": "5430563c",
          "published_at": "2026-09-25T04:53:43.989749Z",
          "score": 0.6055363321799307,
          "model": "gpt-6-sol",
          "runner": "benchmark_agent",
          "config": "openai-gpt-6-sol-high-provider-adapter",
          "verified_publish": true,
          "levels_completed": 1,
          "actions": 107,
          "resets": 2,
          "state": "GAME_OVER",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            16.955017301038062,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            17,
            90,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            7,
            18,
            44,
            61,
            131,
            34,
            152
          ]
        },
        {
          "session_id": "e1193941-aeec-4d34-8361-757610eec9eb",
          "game_id": "vc33",
          "game_version": "5430563c",
          "published_at": "2026-09-25T04:23:59.866283Z",
          "score": 0,
          "model": "gpt-6-sol",
          "runner": "benchmark_agent",
          "config": "openai-gpt-6-sol-high",
          "verified_publish": true,
          "levels_completed": 0,
          "actions": 35,
          "resets": 0,
          "state": "NOT_FINISHED",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            35,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            7,
            18,
            44,
            61,
            131,
            34,
            152
          ]
        },
        {
          "session_id": "2166b983-6148-4c2e-b757-66e92bce2cd9",
          "game_id": "vc33",
          "game_version": "5430563c",
          "published_at": "2026-09-25T03:54:39.520361Z",
          "score": 0,
          "model": "gpt-6-sol",
          "runner": "benchmark_agent",
          "config": "openai-gpt-6-sol-xhigh",
          "verified_publish": true,
          "levels_completed": 0,
          "actions": 35,
          "resets": 1,
          "state": "NOT_FINISHED",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            35,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            7,
            18,
            44,
            61,
            131,
            34,
            152
          ]
        },
        {
          "session_id": "e6deca8c-b7c3-46a0-845c-972663547133",
          "game_id": "vc33",
          "game_version": "5430563c",
          "published_at": "2026-09-21T17:57:13.687011Z",
          "score": 21.428571428571427,
          "model": "gemini-3.8-flash",
          "runner": "benchmark_agent",
          "config": "google-gemini-3-8-flash-low-provider-adapter",
          "verified_publish": true,
          "levels_completed": 3,
          "actions": 352,
          "resets": 5,
          "state": "GAME_OVER",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            100,
            115,
            115,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            7,
            12,
            28,
            305,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            7,
            18,
            44,
            61,
            131,
            34,
            152
          ]
        },
        {
          "session_id": "f0d2965a-20a7-4de1-8338-797300d7f115",
          "game_id": "vc33",
          "game_version": "5430563c",
          "published_at": "2026-09-20T11:14:45.431285Z",
          "score": 0,
          "model": "gemini-3.8-flash",
          "runner": "benchmark_agent",
          "config": "google-gemini-3-8-flash-medium-provider-adapter",
          "verified_publish": true,
          "levels_completed": 0,
          "actions": 35,
          "resets": 2,
          "state": "NOT_FINISHED",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            35,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            7,
            18,
            44,
            61,
            131,
            34,
            152
          ]
        },
        {
          "session_id": "773c394b-78f9-4235-92bb-36053b234e15",
          "game_id": "vc33",
          "game_version": "5430563c",
          "published_at": "2026-09-18T12:40:47.353586Z",
          "score": 100,
          "model": "gemini-3.8-flash",
          "runner": "benchmark_agent",
          "config": "google-gemini-3-8-flash-high-provider-adapter",
          "verified_publish": true,
          "levels_completed": 7,
          "actions": 370,
          "resets": 1,
          "state": "WIN",
          "completed": true,
          "number_of_levels": 0,
          "level_scores": [
            100,
            115,
            115,
            66.15111111111112,
            73.30941091033363,
            115,
            115
          ],
          "level_actions": [
            7,
            11,
            28,
            75,
            153,
            28,
            68
          ],
          "level_baseline_actions": [
            7,
            18,
            44,
            61,
            131,
            34,
            152
          ]
        },
        {
          "session_id": "54bc3aa2-a78b-4e1d-92f3-43f0789a92d3",
          "game_id": "vc33",
          "game_version": "5430563c",
          "published_at": "2026-09-16T19:20:11.629277Z",
          "score": 0,
          "model": "gpt-6-luna",
          "runner": "benchmark_agent",
          "config": "openai-gpt-6-luna-none",
          "verified_publish": true,
          "levels_completed": 0,
          "actions": 35,
          "resets": 2,
          "state": "NOT_FINISHED",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            35,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            7,
            18,
            44,
            61,
            131,
            34,
            152
          ]
        },
        {
          "session_id": "655e1a75-9274-45c3-8d74-8a3db7529358",
          "game_id": "vc33",
          "game_version": "5430563c",
          "published_at": "2026-09-16T19:19:53.398473Z",
          "score": 0,
          "model": "gpt-6-luna",
          "runner": "benchmark_agent",
          "config": "openai-gpt-6-luna-low",
          "verified_publish": true,
          "levels_completed": 0,
          "actions": 35,
          "resets": 1,
          "state": "NOT_FINISHED",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            35,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            7,
            18,
            44,
            61,
            131,
            34,
            152
          ]
        },
        {
          "session_id": "e4d0b106-f627-4a0e-881d-625cb861f7cc",
          "game_id": "vc33",
          "game_version": "5430563c",
          "published_at": "2026-09-16T19:19:03.928587Z",
          "score": 0,
          "model": "gpt-6-luna",
          "runner": "benchmark_agent",
          "config": "openai-gpt-6-luna-low-provider-adapter",
          "verified_publish": true,
          "levels_completed": 0,
          "actions": 35,
          "resets": 0,
          "state": "NOT_FINISHED",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            35,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            7,
            18,
            44,
            61,
            131,
            34,
            152
          ]
        },
        {
          "session_id": "a792f6a5-d014-40b8-a72f-62bbc9b7c11c",
          "game_id": "vc33",
          "game_version": "5430563c",
          "published_at": "2026-09-16T17:34:09.607106Z",
          "score": 0,
          "model": "gpt-6-luna",
          "runner": "benchmark_agent",
          "config": "openai-gpt-6-luna-none-provider-adapter",
          "verified_publish": true,
          "levels_completed": 0,
          "actions": 35,
          "resets": 0,
          "state": "NOT_FINISHED",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            35,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            7,
            18,
            44,
            61,
            131,
            34,
            152
          ]
        },
        {
          "session_id": "9306157c-8843-417a-bb30-2cf5a25431b8",
          "game_id": "vc33",
          "game_version": "5430563c",
          "published_at": "2026-09-16T16:42:01.842661Z",
          "score": 21.428571428571427,
          "model": "gpt-6-1-sol",
          "runner": "benchmark_agent",
          "config": "openai-gpt-6-1-sol-high",
          "verified_publish": true,
          "levels_completed": 3,
          "actions": 363,
          "resets": 7,
          "state": "GAME_OVER",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            40.49586776859504,
            115,
            115,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            11,
            14,
            33,
            305,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            7,
            18,
            44,
            61,
            131,
            34,
            152
          ]
        },
        {
          "session_id": "341e58e4-2a66-45d2-99c2-676008fe699d",
          "game_id": "vc33",
          "game_version": "5430563c",
          "published_at": "2026-09-16T12:22:12.555208Z",
          "score": 54.36440632834496,
          "model": "gpt-6-1-sol",
          "runner": "benchmark_agent",
          "config": "openai-gpt-6-1-sol-max",
          "verified_publish": true,
          "levels_completed": 6,
          "actions": 651,
          "resets": 7,
          "state": "NOT_FINISHED",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            60.49382716049383,
            115,
            115,
            115,
            79.41598408070712,
            4.938271604938271,
            0
          ],
          "level_actions": [
            9,
            11,
            26,
            25,
            147,
            153,
            280
          ],
          "level_baseline_actions": [
            7,
            18,
            44,
            61,
            131,
            34,
            152
          ]
        },
        {
          "session_id": "7c686394-7813-4d5a-8f72-7dda4cee327e",
          "game_id": "vc33",
          "game_version": "5430563c",
          "published_at": "2026-09-16T09:53:58.169901Z",
          "score": 51.51143308185469,
          "model": "gpt-6-1-sol",
          "runner": "benchmark_agent",
          "config": "openai-gpt-6-1-sol-xhigh",
          "verified_publish": true,
          "levels_completed": 6,
          "actions": 878,
          "resets": 9,
          "state": "NOT_FINISHED",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            60.49382716049383,
            115,
            115,
            115,
            28.12863675851104,
            34.363852556480374,
            0
          ],
          "level_actions": [
            9,
            11,
            24,
            49,
            247,
            58,
            480
          ],
          "level_baseline_actions": [
            7,
            18,
            44,
            61,
            131,
            34,
            152
          ]
        },
        {
          "session_id": "503d5c9d-aeaa-4ea8-a5ed-e47d386a854f",
          "game_id": "vc33",
          "game_version": "5430563c",
          "published_at": "2026-09-16T09:00:11.507759Z",
          "score": 0,
          "model": "gpt-6-luna",
          "runner": "benchmark_agent",
          "config": "openai-gpt-6-luna-medium-provider-adapter",
          "verified_publish": true,
          "levels_completed": 0,
          "actions": 35,
          "resets": 1,
          "state": "NOT_FINISHED",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            35,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            7,
            18,
            44,
            61,
            131,
            34,
            152
          ]
        },
        {
          "session_id": "87d37885-e712-4527-956a-07616afb6c49",
          "game_id": "vc33",
          "game_version": "5430563c",
          "published_at": "2026-09-16T08:48:30.011532Z",
          "score": 0,
          "model": "gpt-6-luna",
          "runner": "benchmark_agent",
          "config": "openai-gpt-6-luna-medium",
          "verified_publish": true,
          "levels_completed": 0,
          "actions": 35,
          "resets": 2,
          "state": "NOT_FINISHED",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            35,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            7,
            18,
            44,
            61,
            131,
            34,
            152
          ]
        },
        {
          "session_id": "6dc6ba9c-aec6-460f-bc5e-0c3b7c63298e",
          "game_id": "vc33",
          "game_version": "5430563c",
          "published_at": "2026-09-16T08:29:56.718482Z",
          "score": 0,
          "model": "gpt-6-luna",
          "runner": "benchmark_agent",
          "config": "openai-gpt-6-luna-high-provider-adapter",
          "verified_publish": true,
          "levels_completed": 0,
          "actions": 35,
          "resets": 1,
          "state": "NOT_FINISHED",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            35,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            7,
            18,
            44,
            61,
            131,
            34,
            152
          ]
        },
        {
          "session_id": "84620fd7-e116-4ad0-85af-8aa94f6d6ba1",
          "game_id": "vc33",
          "game_version": "5430563c",
          "published_at": "2026-09-16T08:21:33.677492Z",
          "score": 0,
          "model": "gpt-6-luna",
          "runner": "benchmark_agent",
          "config": "openai-gpt-6-luna-high",
          "verified_publish": true,
          "levels_completed": 0,
          "actions": 35,
          "resets": 2,
          "state": "NOT_FINISHED",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            35,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            7,
            18,
            44,
            61,
            131,
            34,
            152
          ]
        },
        {
          "session_id": "25c75fdf-c0b9-4ab9-9af0-23df11876918",
          "game_id": "vc33",
          "game_version": "5430563c",
          "published_at": "2026-09-16T07:40:38.023239Z",
          "score": 0,
          "model": "gpt-6-luna",
          "runner": "benchmark_agent",
          "config": "openai-gpt-6-luna-xhigh-provider-adapter",
          "verified_publish": true,
          "levels_completed": 0,
          "actions": 35,
          "resets": 1,
          "state": "NOT_FINISHED",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            35,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            7,
            18,
            44,
            61,
            131,
            34,
            152
          ]
        },
        {
          "session_id": "789d703d-9608-46bc-a256-c14a71ee61aa",
          "game_id": "vc33",
          "game_version": "5430563c",
          "published_at": "2026-09-16T07:20:29.777931Z",
          "score": 16.182428148774303,
          "model": "gpt-6-1-sol",
          "runner": "benchmark_agent",
          "config": "openai-gpt-6-1-sol-low",
          "verified_publish": true,
          "levels_completed": 3,
          "actions": 390,
          "resets": 6,
          "state": "GAME_OVER",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            12.249999999999998,
            47.928994082840234,
            115,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            20,
            26,
            39,
            305,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            7,
            18,
            44,
            61,
            131,
            34,
            152
          ]
        },
        {
          "session_id": "1f11c937-a16d-485a-aa7a-6acc07e5c5e4",
          "game_id": "vc33",
          "game_version": "5430563c",
          "published_at": "2026-09-16T07:06:52.400916Z",
          "score": 0,
          "model": "gpt-6-luna",
          "runner": "benchmark_agent",
          "config": "openai-gpt-6-luna-xhigh",
          "verified_publish": true,
          "levels_completed": 0,
          "actions": 35,
          "resets": 1,
          "state": "NOT_FINISHED",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            35,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            7,
            18,
            44,
            61,
            131,
            34,
            152
          ]
        },
        {
          "session_id": "e304c225-7693-49c4-a5cf-f16a07424fbc",
          "game_id": "vc33",
          "game_version": "5430563c",
          "published_at": "2026-09-16T06:10:21.908381Z",
          "score": 0,
          "model": "gpt-6-luna",
          "runner": "benchmark_agent",
          "config": "openai-gpt-6-luna-max",
          "verified_publish": true,
          "levels_completed": 0,
          "actions": 35,
          "resets": 2,
          "state": "NOT_FINISHED",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            35,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            7,
            18,
            44,
            61,
            131,
            34,
            152
          ]
        },
        {
          "session_id": "c3c36df5-ba98-47d1-be93-3f2a6a9b91af",
          "game_id": "vc33",
          "game_version": "5430563c",
          "published_at": "2026-09-16T06:04:04.043521Z",
          "score": 17.889046125117634,
          "model": "gpt-6-1-sol",
          "runner": "benchmark_agent",
          "config": "openai-gpt-6-1-sol-medium",
          "verified_publish": true,
          "levels_completed": 3,
          "actions": 382,
          "resets": 7,
          "state": "GAME_OVER",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            28.994082840236686,
            115,
            80.63306955435236,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            13,
            15,
            49,
            305,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            7,
            18,
            44,
            61,
            131,
            34,
            152
          ]
        },
        {
          "session_id": "3ebcc8d3-3725-42c6-a84c-bfe5170b01e9",
          "game_id": "vc33",
          "game_version": "5430563c",
          "published_at": "2026-09-16T05:03:47.994124Z",
          "score": 100,
          "model": "gpt-6-1-sol",
          "runner": "benchmark_agent",
          "config": "openai-gpt-6-1-sol-high-provider-adapter",
          "verified_publish": true,
          "levels_completed": 7,
          "actions": 312,
          "resets": 1,
          "state": "WIN",
          "completed": true,
          "number_of_levels": 0,
          "level_scores": [
            48.99999999999999,
            115,
            115,
            115,
            115,
            115,
            98.69708231876629
          ],
          "level_actions": [
            10,
            10,
            26,
            26,
            65,
            22,
            153
          ],
          "level_baseline_actions": [
            7,
            18,
            44,
            61,
            131,
            34,
            152
          ]
        },
        {
          "session_id": "f9ee3287-b6a1-4747-bc27-25d2aa327404",
          "game_id": "vc33",
          "game_version": "5430563c",
          "published_at": "2026-09-16T03:38:41.731071Z",
          "score": 100,
          "model": "gpt-6-1-sol",
          "runner": "benchmark_agent",
          "config": "openai-gpt-6-1-sol-xhigh-provider-adapter",
          "verified_publish": true,
          "levels_completed": 7,
          "actions": 187,
          "resets": 0,
          "state": "WIN",
          "completed": true,
          "number_of_levels": 0,
          "level_scores": [
            60.49382716049383,
            115,
            115,
            115,
            115,
            115,
            115
          ],
          "level_actions": [
            9,
            9,
            25,
            24,
            51,
            20,
            49
          ],
          "level_baseline_actions": [
            7,
            18,
            44,
            61,
            131,
            34,
            152
          ]
        },
        {
          "session_id": "055ade2b-7f65-45bc-b119-9c096c85d278",
          "game_id": "vc33",
          "game_version": "5430563c",
          "published_at": "2026-09-16T01:45:39.656431Z",
          "score": 0,
          "model": "gpt-6-luna",
          "runner": "benchmark_agent",
          "config": "openai-gpt-6-luna-max-provider-adapter",
          "verified_publish": true,
          "levels_completed": 0,
          "actions": 35,
          "resets": 1,
          "state": "NOT_FINISHED",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            35,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            7,
            18,
            44,
            61,
            131,
            34,
            152
          ]
        },
        {
          "session_id": "cdbc5529-3bcc-49d1-b0ef-aadbc3c83e67",
          "game_id": "vc33",
          "game_version": "5430563c",
          "published_at": "2026-09-16T01:40:20.601543Z",
          "score": 100,
          "model": "gpt-6-1-sol",
          "runner": "benchmark_agent",
          "config": "openai-gpt-6-1-sol-max-provider-adapter",
          "verified_publish": true,
          "levels_completed": 7,
          "actions": 259,
          "resets": 0,
          "state": "WIN",
          "completed": true,
          "number_of_levels": 0,
          "level_scores": [
            13.573407202216064,
            115,
            115,
            115,
            115,
            115,
            115
          ],
          "level_actions": [
            19,
            15,
            25,
            27,
            104,
            20,
            49
          ],
          "level_baseline_actions": [
            7,
            18,
            44,
            61,
            131,
            34,
            152
          ]
        },
        {
          "session_id": "3a4b5eed-919b-408b-9ba9-cf979cfa8483",
          "game_id": "vc33",
          "game_version": "5430563c",
          "published_at": "2026-09-15T21:58:14.566735Z",
          "score": 14.77777777777778,
          "model": "gemini-3.8-flash",
          "runner": "benchmark_agent",
          "config": "google-gemini-3-8-flash-medium",
          "verified_publish": true,
          "levels_completed": 3,
          "actions": 393,
          "resets": 21,
          "state": "NOT_FINISHED",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            21.77777777777778,
            100,
            64.00000000000001,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            15,
            18,
            55,
            305,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            7,
            18,
            44,
            61,
            131,
            34,
            152
          ]
        },
        {
          "session_id": "811ff597-f339-453d-9a09-83fdde5427e5",
          "game_id": "vc33",
          "game_version": "5430563c",
          "published_at": "2026-09-12T14:24:50.128527Z",
          "score": 10.395503537896857,
          "model": "gemini-3.8-flash",
          "runner": "benchmark_agent",
          "config": "google-gemini-3-8-flash-low",
          "verified_publish": true,
          "levels_completed": 3,
          "actions": 500,
          "resets": 15,
          "state": "GAME_OVER",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            40.49586776859504,
            115,
            6.859410430839002,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            11,
            16,
            168,
            305,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            7,
            18,
            44,
            61,
            131,
            34,
            152
          ]
        },
        {
          "session_id": "ab3d94e4-93b9-4b6a-abbf-931439d11286",
          "game_id": "vc33",
          "game_version": "5430563c",
          "published_at": "2026-09-11T02:09:55.757292Z",
          "score": 21.219308035714285,
          "model": "gemini-3.8-flash",
          "runner": "benchmark_agent",
          "config": "google-gemini-3-8-flash-high",
          "verified_publish": true,
          "levels_completed": 3,
          "actions": 367,
          "resets": 13,
          "state": "GAME_OVER",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            19.140625,
            115,
            115,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            16,
            13,
            33,
            305,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            7,
            18,
            44,
            61,
            131,
            34,
            152
          ]
        },
        {
          "session_id": "2f796865-ac12-40e4-a713-d4e9dd343a6f",
          "game_id": "vc33",
          "game_version": "5430563c",
          "published_at": "2026-08-11T08:17:47.242775Z",
          "score": 0.3968253968253968,
          "model": "grok-4.6",
          "runner": "benchmark_agent",
          "config": "xai-grok-4-6-xhigh",
          "verified_publish": true,
          "levels_completed": 1,
          "actions": 111,
          "resets": 1,
          "state": "GAME_OVER",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            11.11111111111111,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            21,
            90,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            7,
            18,
            44,
            61,
            131,
            34,
            152
          ]
        },
        {
          "session_id": "ce05e7b4-5428-4142-87a9-e805102c0b45",
          "game_id": "vc33",
          "game_version": "5430563c",
          "published_at": "2026-07-21T10:40:20.443737Z",
          "score": 98.81634621832133,
          "model": "claude-opus-5",
          "runner": "benchmark_agent",
          "config": "anthropic-claude-opus-5-high",
          "verified_publish": true,
          "levels_completed": 7,
          "actions": 408,
          "resets": 2,
          "state": "WIN",
          "completed": true,
          "number_of_levels": 0,
          "level_scores": [
            76.5625,
            115,
            115,
            24.595148390508292,
            115,
            106.1524334251607,
            115
          ],
          "level_actions": [
            8,
            8,
            34,
            123,
            118,
            33,
            84
          ],
          "level_baseline_actions": [
            7,
            18,
            44,
            61,
            131,
            34,
            152
          ]
        },
        {
          "session_id": "0a03249d-3422-447a-81df-f6ca4be65c67",
          "game_id": "vc33",
          "game_version": "5430563c",
          "published_at": "2026-06-24T10:51:09.567021Z",
          "score": 0,
          "model": "gpt-5-6-luna",
          "runner": "benchmark_agent",
          "config": "openai-gpt-5-6-luna-low",
          "verified_publish": true,
          "levels_completed": 0,
          "actions": 35,
          "resets": 0,
          "state": "NOT_FINISHED",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            35,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            7,
            18,
            44,
            61,
            131,
            34,
            152
          ]
        },
        {
          "session_id": "048f6957-5709-4c84-8692-b7a897f10ab3",
          "game_id": "vc33",
          "game_version": "5430563c",
          "published_at": "2026-06-24T10:29:34.577035Z",
          "score": 0,
          "model": "gpt-5-6-luna",
          "runner": "benchmark_agent",
          "config": "openai-gpt-5-6-luna-medium",
          "verified_publish": true,
          "levels_completed": 0,
          "actions": 35,
          "resets": 0,
          "state": "NOT_FINISHED",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            35,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            7,
            18,
            44,
            61,
            131,
            34,
            152
          ]
        },
        {
          "session_id": "a00a85bf-d218-454b-bb50-00558700db01",
          "game_id": "vc33",
          "game_version": "5430563c",
          "published_at": "2026-06-24T03:38:44.128695Z",
          "score": 0,
          "model": "gpt-5-6-luna",
          "runner": "benchmark_agent",
          "config": "openai-gpt-5-6-luna-high",
          "verified_publish": true,
          "levels_completed": 0,
          "actions": 35,
          "resets": 3,
          "state": "NOT_FINISHED",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            35,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            7,
            18,
            44,
            61,
            131,
            34,
            152
          ]
        },
        {
          "session_id": "f5233b7f-155c-46b6-ac84-5c7242c9d627",
          "game_id": "vc33",
          "game_version": "5430563c",
          "published_at": "2026-06-24T01:03:28.203826Z",
          "score": 0,
          "model": "gpt-5-6-luna",
          "runner": "benchmark_agent",
          "config": "openai-gpt-5-6-luna-xhigh",
          "verified_publish": true,
          "levels_completed": 0,
          "actions": 35,
          "resets": 0,
          "state": "NOT_FINISHED",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            35,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            7,
            18,
            44,
            61,
            131,
            34,
            152
          ]
        },
        {
          "session_id": "16e3558f-764e-41f2-af4e-43de6c0fa209",
          "game_id": "vc33",
          "game_version": "5430563c",
          "published_at": "2026-06-23T00:11:15.023502Z",
          "score": 0,
          "model": "gpt-5-6-luna",
          "runner": "benchmark_agent",
          "config": "openai-gpt-5-6-luna-max",
          "verified_publish": true,
          "levels_completed": 0,
          "actions": 35,
          "resets": 3,
          "state": "NOT_FINISHED",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            35,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            7,
            18,
            44,
            61,
            131,
            34,
            152
          ]
        },
        {
          "session_id": "e4485500-8697-42e6-9aae-68fd594b2382",
          "game_id": "vc33",
          "game_version": "5430563c",
          "published_at": "2026-06-20T11:57:41.838618Z",
          "score": 0,
          "model": "gpt-5-6-terra",
          "runner": "benchmark_agent",
          "config": "openai-gpt-5-6-terra-low",
          "verified_publish": true,
          "levels_completed": 0,
          "actions": 35,
          "resets": 0,
          "state": "NOT_FINISHED",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            35,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            7,
            18,
            44,
            61,
            131,
            34,
            152
          ]
        },
        {
          "session_id": "8363458e-604c-4b55-86f1-3b26e1a551ba",
          "game_id": "vc33",
          "game_version": "5430563c",
          "published_at": "2026-06-20T09:21:38.584346Z",
          "score": 0,
          "model": "gpt-5-6-terra",
          "runner": "benchmark_agent",
          "config": "openai-gpt-5-6-terra-medium",
          "verified_publish": true,
          "levels_completed": 0,
          "actions": 35,
          "resets": 0,
          "state": "NOT_FINISHED",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            35,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            7,
            18,
            44,
            61,
            131,
            34,
            152
          ]
        },
        {
          "session_id": "9371fc79-8411-4937-add2-6132cadd70ee",
          "game_id": "vc33",
          "game_version": "5430563c",
          "published_at": "2026-06-20T08:34:38.76273Z",
          "score": 0,
          "model": "gpt-5-6-terra",
          "runner": "benchmark_agent",
          "config": "openai-gpt-5-6-terra-high",
          "verified_publish": true,
          "levels_completed": 0,
          "actions": 35,
          "resets": 0,
          "state": "NOT_FINISHED",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            35,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            7,
            18,
            44,
            61,
            131,
            34,
            152
          ]
        },
        {
          "session_id": "7532cdff-0e9d-432f-8884-5190c746ffa5",
          "game_id": "vc33",
          "game_version": "5430563c",
          "published_at": "2026-06-20T02:10:04.271427Z",
          "score": 0,
          "model": "gpt-5-6-terra",
          "runner": "benchmark_agent",
          "config": "openai-gpt-5-6-terra-xhigh",
          "verified_publish": true,
          "levels_completed": 0,
          "actions": 35,
          "resets": 1,
          "state": "NOT_FINISHED",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            35,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            7,
            18,
            44,
            61,
            131,
            34,
            152
          ]
        },
        {
          "session_id": "bfa09944-be3e-4588-9e08-ececc004e766",
          "game_id": "vc33",
          "game_version": "5430563c",
          "published_at": "2026-06-18T20:29:50.966328Z",
          "score": 0,
          "model": "gpt-5-6-terra",
          "runner": "benchmark_agent",
          "config": "openai-gpt-5-6-terra-max",
          "verified_publish": true,
          "levels_completed": 0,
          "actions": 35,
          "resets": 7,
          "state": "NOT_FINISHED",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            35,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            7,
            18,
            44,
            61,
            131,
            34,
            152
          ]
        },
        {
          "session_id": "8239e65e-dc8d-470d-ab49-6d37618f00ea",
          "game_id": "vc33",
          "game_version": "5430563c",
          "published_at": "2026-06-13T08:31:42.449999Z",
          "score": 21.428571428571427,
          "model": "gpt-5-6-sol",
          "runner": "benchmark_agent",
          "config": "openai-gpt-5-6-sol-max",
          "verified_publish": true,
          "levels_completed": 3,
          "actions": 352,
          "resets": 8,
          "state": "GAME_OVER",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            100,
            115,
            115,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            7,
            11,
            29,
            305,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            7,
            18,
            44,
            61,
            131,
            34,
            152
          ]
        },
        {
          "session_id": "c604a83e-2fe5-4e86-b95c-0934deca7e91",
          "game_id": "vc33",
          "game_version": "5430563c",
          "published_at": "2026-06-12T07:22:37.346719Z",
          "score": 16.366354607375403,
          "model": "gpt-5-6-sol",
          "runner": "benchmark_agent",
          "config": "openai-gpt-5-6-sol-high",
          "verified_publish": true,
          "levels_completed": 3,
          "actions": 386,
          "resets": 16,
          "state": "GAME_OVER",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            21.77777777777778,
            81,
            91.49338374291116,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            15,
            20,
            46,
            305,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            7,
            18,
            44,
            61,
            131,
            34,
            152
          ]
        },
        {
          "session_id": "d199a6b5-9709-4985-a373-9fff94c6ba4b",
          "game_id": "vc33",
          "game_version": "5430563c",
          "published_at": "2026-06-12T00:31:25.19315Z",
          "score": 11.505817903855954,
          "model": "gpt-5-6-sol",
          "runner": "benchmark_agent",
          "config": "openai-gpt-5-6-sol-xhigh",
          "verified_publish": true,
          "levels_completed": 3,
          "actions": 348,
          "resets": 27,
          "state": "NOT_FINISHED",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            48.99999999999999,
            115,
            14.387633769322234,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            10,
            9,
            116,
            213,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            7,
            18,
            44,
            61,
            131,
            34,
            152
          ]
        },
        {
          "session_id": "38c7a651-d71e-4337-9efb-1f0af97fd06e",
          "game_id": "vc33",
          "game_version": "5430563c",
          "published_at": "2026-06-10T08:41:05.542137Z",
          "score": 0,
          "model": "gpt-5-6-sol",
          "runner": "benchmark_agent",
          "config": "openai-gpt-5-6-sol-medium",
          "verified_publish": true,
          "levels_completed": 0,
          "actions": 35,
          "resets": 0,
          "state": "NOT_FINISHED",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            35,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            7,
            18,
            44,
            61,
            131,
            34,
            152
          ]
        },
        {
          "session_id": "f57ab031-8b93-44b1-a36a-0b435805f4d3",
          "game_id": "vc33",
          "game_version": "5430563c",
          "published_at": "2026-06-09T22:15:43.543447Z",
          "score": 0,
          "model": "gpt-5-6-sol",
          "runner": "benchmark_agent",
          "config": "openai-gpt-5-6-sol-low",
          "verified_publish": true,
          "levels_completed": 0,
          "actions": 35,
          "resets": 0,
          "state": "NOT_FINISHED",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            35,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            7,
            18,
            44,
            61,
            131,
            34,
            152
          ]
        },
        {
          "session_id": "6cec87d3-563e-47eb-9a2d-1a66a9e85479",
          "game_id": "vc33",
          "game_version": "5430563c",
          "published_at": "2026-05-29T22:25:56.212762Z",
          "score": 9.964049233588838,
          "model": "claude-opus-4-8",
          "runner": "benchmark_agent",
          "config": "anthropic-opus-4-8-high",
          "verified_publish": true,
          "levels_completed": 3,
          "actions": 388,
          "resets": 10,
          "state": "GAME_OVER",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            34.027777777777786,
            115,
            4.988533587569894,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            12,
            15,
            197,
            164,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            7,
            18,
            44,
            61,
            131,
            34,
            152
          ]
        },
        {
          "session_id": "c9c41f63-1885-4287-9a2e-29f24bb92ce3",
          "game_id": "vc33",
          "game_version": "5430563c",
          "published_at": "2026-05-01T16:15:59.418513Z",
          "score": 0,
          "model": "gpt-5.5-2026-04-23",
          "runner": "benchmark_agent",
          "config": "openai-gpt-5-5-2026-04-23-high",
          "verified_publish": true,
          "levels_completed": 0,
          "actions": 35,
          "resets": 0,
          "state": "NOT_FINISHED",
          "completed": false,
          "number_of_levels": 7,
          "level_scores": [
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            35,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            7,
            18,
            44,
            61,
            131,
            34,
            152
          ]
        },
        {
          "session_id": "e29db330-d6f5-4bd8-ae22-66b4b8c80172",
          "game_id": "vc33",
          "game_version": "5430563c",
          "published_at": "2026-05-01T05:06:43.175312Z",
          "score": 0,
          "model": "claude-opus-4-7",
          "runner": "benchmark_agent",
          "config": "anthropic-opus-4-7-high",
          "verified_publish": true,
          "levels_completed": 0,
          "actions": 35,
          "resets": 2,
          "state": "NOT_FINISHED",
          "completed": false,
          "number_of_levels": 7,
          "level_scores": [
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            35,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            7,
            18,
            44,
            61,
            131,
            34,
            152
          ]
        },
        {
          "session_id": "bb79ce0a-1373-4aaa-957b-75de1d1dfa87",
          "game_id": "vc33",
          "game_version": "5430563c",
          "published_at": "2026-04-14T17:44:05.024594Z",
          "score": 0,
          "model": "grok-4.20-beta-0309-reasoning",
          "runner": "benchmark_agent",
          "config": "xai-grok-4-20-beta-0309-reasoning",
          "verified_publish": true,
          "levels_completed": 0,
          "actions": 35,
          "resets": 1,
          "state": "NOT_FINISHED",
          "completed": false,
          "number_of_levels": 7,
          "level_scores": [
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            35,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            7,
            18,
            44,
            61,
            131,
            34,
            152
          ]
        },
        {
          "session_id": "6639a825-4862-410b-bf57-aedd481c9671",
          "game_id": "vc33",
          "game_version": "5430563c",
          "published_at": "2026-04-14T17:44:04.706737Z",
          "score": 0,
          "model": "gpt-5.4-2026-03-05",
          "runner": "benchmark_agent",
          "config": "openai-gpt-5-4-2026-03-05-high",
          "verified_publish": true,
          "levels_completed": 0,
          "actions": 35,
          "resets": 0,
          "state": "NOT_FINISHED",
          "completed": false,
          "number_of_levels": 7,
          "level_scores": [
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            35,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            7,
            18,
            44,
            61,
            131,
            34,
            152
          ]
        },
        {
          "session_id": "23b364ed-4e2d-487e-8ad6-fcd71ea16f4f",
          "game_id": "vc33",
          "game_version": "5430563c",
          "published_at": "2026-04-14T17:44:04.070943Z",
          "score": 0,
          "model": "gemini-3.1-pro-preview",
          "runner": "benchmark_agent",
          "config": "google-gemini-3-1-pro-preview",
          "verified_publish": true,
          "levels_completed": 0,
          "actions": 35,
          "resets": 3,
          "state": "NOT_FINISHED",
          "completed": false,
          "number_of_levels": 7,
          "level_scores": [
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            35,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            7,
            18,
            44,
            61,
            131,
            34,
            152
          ]
        },
        {
          "session_id": "24e86490-c706-4f70-95ec-2d8a45df6ca6",
          "game_id": "vc33",
          "game_version": "5430563c",
          "published_at": "2026-04-14T17:44:03.777826Z",
          "score": 0,
          "model": "claude-opus-4-6",
          "runner": "benchmark_agent",
          "config": "anthropic-opus-4-6-max-effort",
          "verified_publish": true,
          "levels_completed": 0,
          "actions": 35,
          "resets": 1,
          "state": "NOT_FINISHED",
          "completed": false,
          "number_of_levels": 7,
          "level_scores": [
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            35,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            7,
            18,
            44,
            61,
            131,
            34,
            152
          ]
        }
      ]
    },
    {
      "game": "cn04",
      "source": "https://arcprize.org/tasks/cn04",
      "runs": [
        {
          "session_id": "331285c6-5990-4061-b5b7-ad8d49bbc88f",
          "game_id": "cn04",
          "game_version": "2fe56bfb",
          "published_at": "2026-09-25T20:40:49.872116Z",
          "score": 43.01092565049768,
          "model": "gpt-6-sol",
          "runner": "benchmark_agent",
          "config": "openai-gpt-6-sol-medium-provider-adapter",
          "verified_publish": true,
          "levels_completed": 4,
          "actions": 1435,
          "resets": 9,
          "state": "GAME_OVER",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            87.51300728407907,
            92.98469387755102,
            56.582347873756746,
            115,
            0,
            0
          ],
          "level_actions": [
            31,
            56,
            113,
            195,
            1040,
            0
          ],
          "level_baseline_actions": [
            29,
            54,
            85,
            300,
            208,
            113
          ]
        },
        {
          "session_id": "8cb667c4-2cf9-4f61-90f5-c542ef741780",
          "game_id": "cn04",
          "game_version": "2fe56bfb",
          "published_at": "2026-09-25T15:30:22.519812Z",
          "score": 1.2030848201600912,
          "model": "gpt-6-sol",
          "runner": "benchmark_agent",
          "config": "openai-gpt-6-sol-max",
          "verified_publish": true,
          "levels_completed": 2,
          "actions": 336,
          "resets": 5,
          "state": "GAME_OVER",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            8.938250611117017,
            8.16326530612245,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            97,
            189,
            50,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            29,
            54,
            85,
            300,
            208,
            113
          ]
        },
        {
          "session_id": "cbc01692-ca11-4364-9d63-ad951cbba9fd",
          "game_id": "cn04",
          "game_version": "2fe56bfb",
          "published_at": "2026-09-25T12:04:31.681415Z",
          "score": 0,
          "model": "gpt-6-sol",
          "runner": "benchmark_agent",
          "config": "openai-gpt-6-sol-none",
          "verified_publish": true,
          "levels_completed": 0,
          "actions": 145,
          "resets": 1,
          "state": "GAME_OVER",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            145,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            29,
            54,
            85,
            300,
            208,
            113
          ]
        },
        {
          "session_id": "a3321fec-4db6-40df-8c5e-e870e7fdbbe6",
          "game_id": "cn04",
          "game_version": "2fe56bfb",
          "published_at": "2026-09-25T09:09:22.159775Z",
          "score": 47.61904761904761,
          "model": "gpt-6-sol",
          "runner": "benchmark_agent",
          "config": "openai-gpt-6-sol-high-provider-adapter",
          "verified_publish": true,
          "levels_completed": 4,
          "actions": 1304,
          "resets": 9,
          "state": "GAME_OVER",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            115,
            115,
            115,
            115,
            0,
            0
          ],
          "level_actions": [
            17,
            50,
            42,
            155,
            1040,
            0
          ],
          "level_baseline_actions": [
            29,
            54,
            85,
            300,
            208,
            113
          ]
        },
        {
          "session_id": "4c88afb2-4b9e-4b44-aad5-2e6d6dc22868",
          "game_id": "cn04",
          "game_version": "2fe56bfb",
          "published_at": "2026-09-25T08:09:31.874934Z",
          "score": 4.761904761904762,
          "model": "gpt-6-sol",
          "runner": "benchmark_agent",
          "config": "openai-gpt-6-sol-xhigh",
          "verified_publish": true,
          "levels_completed": 1,
          "actions": 297,
          "resets": 2,
          "state": "GAME_OVER",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            115,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            27,
            270,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            29,
            54,
            85,
            300,
            208,
            113
          ]
        },
        {
          "session_id": "51d8fc3b-7e6b-4a8a-8080-e402c3581c6f",
          "game_id": "cn04",
          "game_version": "2fe56bfb",
          "published_at": "2026-09-25T06:46:17.087371Z",
          "score": 0.6416859325047115,
          "model": "gpt-6-sol",
          "runner": "benchmark_agent",
          "config": "openai-gpt-6-sol-low-provider-adapter",
          "verified_publish": true,
          "levels_completed": 1,
          "actions": 349,
          "resets": 4,
          "state": "GAME_OVER",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            13.475404582598943,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            79,
            270,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            29,
            54,
            85,
            300,
            208,
            113
          ]
        },
        {
          "session_id": "03a65d3a-6981-43f0-af50-ee847c688bdb",
          "game_id": "cn04",
          "game_version": "2fe56bfb",
          "published_at": "2026-09-25T06:42:58.263831Z",
          "score": 0,
          "model": "gpt-6-sol",
          "runner": "benchmark_agent",
          "config": "openai-gpt-6-sol-none-provider-adapter",
          "verified_publish": true,
          "levels_completed": 0,
          "actions": 145,
          "resets": 4,
          "state": "NOT_FINISHED",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            145,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            29,
            54,
            85,
            300,
            208,
            113
          ]
        },
        {
          "session_id": "af3ded84-b87e-4ce0-8cb9-5e18065ced45",
          "game_id": "cn04",
          "game_version": "2fe56bfb",
          "published_at": "2026-09-25T06:32:46.022367Z",
          "score": 3.67746731383095,
          "model": "gpt-6-sol",
          "runner": "benchmark_agent",
          "config": "openai-gpt-6-sol-high",
          "verified_publish": true,
          "levels_completed": 1,
          "actions": 303,
          "resets": 3,
          "state": "GAME_OVER",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            77.22681359044995,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            33,
            270,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            29,
            54,
            85,
            300,
            208,
            113
          ]
        },
        {
          "session_id": "969505f5-5e86-4b73-866d-cec021619ec9",
          "game_id": "cn04",
          "game_version": "2fe56bfb",
          "published_at": "2026-09-25T06:03:55.179572Z",
          "score": 0,
          "model": "gpt-6-sol",
          "runner": "benchmark_agent",
          "config": "openai-gpt-6-sol-low",
          "verified_publish": true,
          "levels_completed": 0,
          "actions": 145,
          "resets": 1,
          "state": "GAME_OVER",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            145,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            29,
            54,
            85,
            300,
            208,
            113
          ]
        },
        {
          "session_id": "7043b64d-a16b-41e4-a992-d6007feb76e3",
          "game_id": "cn04",
          "game_version": "2fe56bfb",
          "published_at": "2026-09-25T05:44:03.814909Z",
          "score": 0,
          "model": "gpt-6-sol",
          "runner": "benchmark_agent",
          "config": "openai-gpt-6-sol-medium",
          "verified_publish": true,
          "levels_completed": 0,
          "actions": 145,
          "resets": 2,
          "state": "GAME_OVER",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            145,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            29,
            54,
            85,
            300,
            208,
            113
          ]
        },
        {
          "session_id": "9612181a-9a83-4235-8e31-b555b24be0fa",
          "game_id": "cn04",
          "game_version": "2fe56bfb",
          "published_at": "2026-09-25T05:07:25.401359Z",
          "score": 4.761904761904762,
          "model": "gpt-6-sol",
          "runner": "benchmark_agent",
          "config": "openai-gpt-6-sol-xhigh-provider-adapter",
          "verified_publish": true,
          "levels_completed": 1,
          "actions": 293,
          "resets": 3,
          "state": "GAME_OVER",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            115,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            23,
            270,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            29,
            54,
            85,
            300,
            208,
            113
          ]
        },
        {
          "session_id": "aa18dd34-ec86-48c9-a85a-9c0aeeff162d",
          "game_id": "cn04",
          "game_version": "2fe56bfb",
          "published_at": "2026-09-21T11:30:33.185917Z",
          "score": 39.88762261721989,
          "model": "gemini-3.8-flash",
          "runner": "benchmark_agent",
          "config": "google-gemini-3-8-flash-low-provider-adapter",
          "verified_publish": true,
          "levels_completed": 4,
          "actions": 1433,
          "resets": 38,
          "state": "NOT_FINISHED",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            115,
            47.928994082840234,
            55.59402893197907,
            115,
            0,
            0
          ],
          "level_actions": [
            14,
            78,
            114,
            187,
            1040,
            0
          ],
          "level_baseline_actions": [
            29,
            54,
            85,
            300,
            208,
            113
          ]
        },
        {
          "session_id": "c4c98dcc-e9bf-4d82-9c4b-79f02e0cc51c",
          "game_id": "cn04",
          "game_version": "2fe56bfb",
          "published_at": "2026-09-20T03:49:13.722631Z",
          "score": 4.761904761904762,
          "model": "gemini-3.8-flash",
          "runner": "benchmark_agent",
          "config": "google-gemini-3-8-flash-medium-provider-adapter",
          "verified_publish": true,
          "levels_completed": 1,
          "actions": 286,
          "resets": 3,
          "state": "GAME_OVER",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            115,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            16,
            270,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            29,
            54,
            85,
            300,
            208,
            113
          ]
        },
        {
          "session_id": "c92760f4-72a8-4de0-8ddf-c662b340d61c",
          "game_id": "cn04",
          "game_version": "2fe56bfb",
          "published_at": "2026-09-18T02:09:40.017831Z",
          "score": 34.20683703331371,
          "model": "gemini-3.8-flash",
          "runner": "benchmark_agent",
          "config": "google-gemini-3-8-flash-high-provider-adapter",
          "verified_publish": true,
          "levels_completed": 4,
          "actions": 1698,
          "resets": 13,
          "state": "GAME_OVER",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            115,
            115,
            115,
            7.085894424896999,
            0,
            0
          ],
          "level_actions": [
            16,
            31,
            66,
            1127,
            458,
            0
          ],
          "level_baseline_actions": [
            29,
            54,
            85,
            300,
            208,
            113
          ]
        },
        {
          "session_id": "37b2a9ce-dfd3-4635-9bbe-186eb99d44d4",
          "game_id": "cn04",
          "game_version": "2fe56bfb",
          "published_at": "2026-09-16T20:21:07.501554Z",
          "score": 0,
          "model": "gpt-6-luna",
          "runner": "benchmark_agent",
          "config": "openai-gpt-6-luna-low-provider-adapter",
          "verified_publish": true,
          "levels_completed": 0,
          "actions": 145,
          "resets": 1,
          "state": "GAME_OVER",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            145,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            29,
            54,
            85,
            300,
            208,
            113
          ]
        },
        {
          "session_id": "d0f3d3c5-876d-4e3a-93f6-ac22ee89a5af",
          "game_id": "cn04",
          "game_version": "2fe56bfb",
          "published_at": "2026-09-16T19:25:13.089745Z",
          "score": 0,
          "model": "gpt-6-luna",
          "runner": "benchmark_agent",
          "config": "openai-gpt-6-luna-low",
          "verified_publish": true,
          "levels_completed": 0,
          "actions": 145,
          "resets": 1,
          "state": "GAME_OVER",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            145,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            29,
            54,
            85,
            300,
            208,
            113
          ]
        },
        {
          "session_id": "31f3e06f-9841-4d07-94a4-ebb642104160",
          "game_id": "cn04",
          "game_version": "2fe56bfb",
          "published_at": "2026-09-16T19:24:13.539204Z",
          "score": 0,
          "model": "gpt-6-luna",
          "runner": "benchmark_agent",
          "config": "openai-gpt-6-luna-none",
          "verified_publish": true,
          "levels_completed": 0,
          "actions": 145,
          "resets": 2,
          "state": "GAME_OVER",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            145,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            29,
            54,
            85,
            300,
            208,
            113
          ]
        },
        {
          "session_id": "7caaa05a-d61d-472d-8917-208a26232748",
          "game_id": "cn04",
          "game_version": "2fe56bfb",
          "published_at": "2026-09-16T17:41:43.462765Z",
          "score": 0,
          "model": "gpt-6-luna",
          "runner": "benchmark_agent",
          "config": "openai-gpt-6-luna-medium-provider-adapter",
          "verified_publish": true,
          "levels_completed": 0,
          "actions": 145,
          "resets": 1,
          "state": "GAME_OVER",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            145,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            29,
            54,
            85,
            300,
            208,
            113
          ]
        },
        {
          "session_id": "e310b137-f84c-4348-808c-c233057648ad",
          "game_id": "cn04",
          "game_version": "2fe56bfb",
          "published_at": "2026-09-16T09:15:17.771026Z",
          "score": 0,
          "model": "gpt-6-luna",
          "runner": "benchmark_agent",
          "config": "openai-gpt-6-luna-none-provider-adapter",
          "verified_publish": true,
          "levels_completed": 0,
          "actions": 145,
          "resets": 2,
          "state": "GAME_OVER",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            145,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            29,
            54,
            85,
            300,
            208,
            113
          ]
        },
        {
          "session_id": "cb5a05a3-e349-44e3-9957-ea2e37f11816",
          "game_id": "cn04",
          "game_version": "2fe56bfb",
          "published_at": "2026-09-16T09:04:09.281981Z",
          "score": 0,
          "model": "gpt-6-luna",
          "runner": "benchmark_agent",
          "config": "openai-gpt-6-luna-medium",
          "verified_publish": true,
          "levels_completed": 0,
          "actions": 145,
          "resets": 1,
          "state": "GAME_OVER",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            145,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            29,
            54,
            85,
            300,
            208,
            113
          ]
        },
        {
          "session_id": "17e0ec93-41ac-4d16-bc7e-563c40bba8aa",
          "game_id": "cn04",
          "game_version": "2fe56bfb",
          "published_at": "2026-09-16T08:52:30.795345Z",
          "score": 0,
          "model": "gpt-6-luna",
          "runner": "benchmark_agent",
          "config": "openai-gpt-6-luna-high",
          "verified_publish": true,
          "levels_completed": 0,
          "actions": 145,
          "resets": 1,
          "state": "GAME_OVER",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            145,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            29,
            54,
            85,
            300,
            208,
            113
          ]
        },
        {
          "session_id": "c20d85fe-1584-4227-8a5c-bd6237e3b21a",
          "game_id": "cn04",
          "game_version": "2fe56bfb",
          "published_at": "2026-09-16T08:52:20.126988Z",
          "score": 0,
          "model": "gpt-6-luna",
          "runner": "benchmark_agent",
          "config": "openai-gpt-6-luna-high-provider-adapter",
          "verified_publish": true,
          "levels_completed": 0,
          "actions": 145,
          "resets": 2,
          "state": "GAME_OVER",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            145,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            29,
            54,
            85,
            300,
            208,
            113
          ]
        },
        {
          "session_id": "d485f233-9c66-402d-b144-82d098c85a20",
          "game_id": "cn04",
          "game_version": "2fe56bfb",
          "published_at": "2026-09-16T08:51:39.373178Z",
          "score": 58.48501496335063,
          "model": "gpt-6-1-sol",
          "runner": "benchmark_agent",
          "config": "openai-gpt-6-1-sol-high",
          "verified_publish": true,
          "levels_completed": 5,
          "actions": 1225,
          "resets": 10,
          "state": "NOT_FINISHED",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            115,
            115,
            115,
            115,
            15.637062846072663,
            0
          ],
          "level_actions": [
            18,
            33,
            28,
            55,
            526,
            565
          ],
          "level_baseline_actions": [
            29,
            54,
            85,
            300,
            208,
            113
          ]
        },
        {
          "session_id": "1fa61313-3410-43c8-a4e6-d1217c5c2efe",
          "game_id": "cn04",
          "game_version": "2fe56bfb",
          "published_at": "2026-09-16T08:39:55.895538Z",
          "score": 0,
          "model": "gpt-6-luna",
          "runner": "benchmark_agent",
          "config": "openai-gpt-6-luna-max",
          "verified_publish": true,
          "levels_completed": 0,
          "actions": 145,
          "resets": 4,
          "state": "NOT_FINISHED",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            145,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            29,
            54,
            85,
            300,
            208,
            113
          ]
        },
        {
          "session_id": "4612c21b-9eb4-4e99-b109-a49905482138",
          "game_id": "cn04",
          "game_version": "2fe56bfb",
          "published_at": "2026-09-16T08:12:30.424902Z",
          "score": 0,
          "model": "gpt-6-luna",
          "runner": "benchmark_agent",
          "config": "openai-gpt-6-luna-xhigh-provider-adapter",
          "verified_publish": true,
          "levels_completed": 0,
          "actions": 145,
          "resets": 1,
          "state": "NOT_FINISHED",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            145,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            29,
            54,
            85,
            300,
            208,
            113
          ]
        },
        {
          "session_id": "3b824760-684c-4cef-8f64-0cf1195411db",
          "game_id": "cn04",
          "game_version": "2fe56bfb",
          "published_at": "2026-09-16T08:03:49.320059Z",
          "score": 38.10130320842645,
          "model": "gpt-6-1-sol",
          "runner": "benchmark_agent",
          "config": "openai-gpt-6-1-sol-medium",
          "verified_publish": true,
          "levels_completed": 4,
          "actions": 1576,
          "resets": 12,
          "state": "GAME_OVER",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            100,
            14.259866008117756,
            115,
            81.65190884018,
            0,
            0
          ],
          "level_actions": [
            29,
            143,
            32,
            332,
            1040,
            0
          ],
          "level_baseline_actions": [
            29,
            54,
            85,
            300,
            208,
            113
          ]
        },
        {
          "session_id": "fb3c1c4e-baec-4921-a7e4-f1a7ad67878d",
          "game_id": "cn04",
          "game_version": "2fe56bfb",
          "published_at": "2026-09-16T07:56:50.563738Z",
          "score": 0,
          "model": "gpt-6-luna",
          "runner": "benchmark_agent",
          "config": "openai-gpt-6-luna-xhigh",
          "verified_publish": true,
          "levels_completed": 0,
          "actions": 145,
          "resets": 3,
          "state": "NOT_FINISHED",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            145,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            29,
            54,
            85,
            300,
            208,
            113
          ]
        },
        {
          "session_id": "ec68f9da-0df5-464f-8be4-146ebf0debe7",
          "game_id": "cn04",
          "game_version": "2fe56bfb",
          "published_at": "2026-09-16T07:19:55.106576Z",
          "score": 1.1504630579609034,
          "model": "gpt-6-1-sol",
          "runner": "benchmark_agent",
          "config": "openai-gpt-6-1-sol-low",
          "verified_publish": true,
          "levels_completed": 1,
          "actions": 329,
          "resets": 2,
          "state": "GAME_OVER",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            24.159724217178972,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            59,
            270,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            29,
            54,
            85,
            300,
            208,
            113
          ]
        },
        {
          "session_id": "d48df2b9-4cfc-4504-843a-d539780bdc68",
          "game_id": "cn04",
          "game_version": "2fe56bfb",
          "published_at": "2026-09-16T05:37:25.887339Z",
          "score": 92.9374024554489,
          "model": "gpt-6-1-sol",
          "runner": "benchmark_agent",
          "config": "openai-gpt-6-1-sol-xhigh",
          "verified_publish": true,
          "levels_completed": 6,
          "actions": 483,
          "resets": 3,
          "state": "WIN",
          "completed": true,
          "number_of_levels": 0,
          "level_scores": [
            115,
            115,
            115,
            115,
            101.95117353190687,
            48.65493065081542
          ],
          "level_actions": [
            14,
            30,
            32,
            39,
            206,
            162
          ],
          "level_baseline_actions": [
            29,
            54,
            85,
            300,
            208,
            113
          ]
        },
        {
          "session_id": "551d7521-dd4f-493e-b0f9-5601f3d0537b",
          "game_id": "cn04",
          "game_version": "2fe56bfb",
          "published_at": "2026-09-16T05:24:42.237756Z",
          "score": 100,
          "model": "gpt-6-1-sol",
          "runner": "benchmark_agent",
          "config": "openai-gpt-6-1-sol-max",
          "verified_publish": true,
          "levels_completed": 6,
          "actions": 241,
          "resets": 0,
          "state": "WIN",
          "completed": true,
          "number_of_levels": 0,
          "level_scores": [
            115,
            115,
            115,
            115,
            115,
            115
          ],
          "level_actions": [
            17,
            48,
            34,
            34,
            61,
            47
          ],
          "level_baseline_actions": [
            29,
            54,
            85,
            300,
            208,
            113
          ]
        },
        {
          "session_id": "090a79c1-549e-4124-a721-8128ed491724",
          "game_id": "cn04",
          "game_version": "2fe56bfb",
          "published_at": "2026-09-16T05:12:18.867023Z",
          "score": 100,
          "model": "gpt-6-1-sol",
          "runner": "benchmark_agent",
          "config": "openai-gpt-6-1-sol-high-provider-adapter",
          "verified_publish": true,
          "levels_completed": 6,
          "actions": 278,
          "resets": 0,
          "state": "WIN",
          "completed": true,
          "number_of_levels": 0,
          "level_scores": [
            115,
            115,
            115,
            115,
            115,
            115
          ],
          "level_actions": [
            14,
            46,
            28,
            38,
            111,
            41
          ],
          "level_baseline_actions": [
            29,
            54,
            85,
            300,
            208,
            113
          ]
        },
        {
          "session_id": "b757663a-8a19-451d-9abc-e10dce7dab33",
          "game_id": "cn04",
          "game_version": "2fe56bfb",
          "published_at": "2026-09-16T03:55:47.307699Z",
          "score": 100,
          "model": "gpt-6-1-sol",
          "runner": "benchmark_agent",
          "config": "openai-gpt-6-1-sol-max-provider-adapter",
          "verified_publish": true,
          "levels_completed": 6,
          "actions": 244,
          "resets": 0,
          "state": "WIN",
          "completed": true,
          "number_of_levels": 0,
          "level_scores": [
            115,
            115,
            115,
            115,
            115,
            115
          ],
          "level_actions": [
            14,
            29,
            32,
            32,
            97,
            40
          ],
          "level_baseline_actions": [
            29,
            54,
            85,
            300,
            208,
            113
          ]
        },
        {
          "session_id": "c153c397-25c5-4dee-8d56-950a1bb4bb03",
          "game_id": "cn04",
          "game_version": "2fe56bfb",
          "published_at": "2026-09-16T03:05:16.565647Z",
          "score": 100,
          "model": "gpt-6-1-sol",
          "runner": "benchmark_agent",
          "config": "openai-gpt-6-1-sol-xhigh-provider-adapter",
          "verified_publish": true,
          "levels_completed": 6,
          "actions": 255,
          "resets": 0,
          "state": "WIN",
          "completed": true,
          "number_of_levels": 0,
          "level_scores": [
            115,
            115,
            115,
            115,
            115,
            115
          ],
          "level_actions": [
            17,
            29,
            25,
            32,
            111,
            41
          ],
          "level_baseline_actions": [
            29,
            54,
            85,
            300,
            208,
            113
          ]
        },
        {
          "session_id": "0bd6297c-048d-4713-b953-f87f949dc0f0",
          "game_id": "cn04",
          "game_version": "2fe56bfb",
          "published_at": "2026-09-15T21:41:50.268888Z",
          "score": 0,
          "model": "gpt-6-luna",
          "runner": "benchmark_agent",
          "config": "openai-gpt-6-luna-max-provider-adapter",
          "verified_publish": true,
          "levels_completed": 0,
          "actions": 145,
          "resets": 2,
          "state": "GAME_OVER",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            145,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            29,
            54,
            85,
            300,
            208,
            113
          ]
        },
        {
          "session_id": "081f3ff1-a95e-47ed-8532-8b2641afdd74",
          "game_id": "cn04",
          "game_version": "2fe56bfb",
          "published_at": "2026-09-12T15:47:27.81841Z",
          "score": 30.29156911488426,
          "model": "gemini-3.8-flash",
          "runner": "benchmark_agent",
          "config": "google-gemini-3-8-flash-high",
          "verified_publish": true,
          "levels_completed": 4,
          "actions": 589,
          "resets": 23,
          "state": "NOT_FINISHED",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            115,
            4.51980903961808,
            17.361111111111114,
            115,
            0,
            0
          ],
          "level_actions": [
            16,
            254,
            204,
            50,
            65,
            0
          ],
          "level_baseline_actions": [
            29,
            54,
            85,
            300,
            208,
            113
          ]
        },
        {
          "session_id": "fe651579-8c28-4323-976f-bb41f705934e",
          "game_id": "cn04",
          "game_version": "2fe56bfb",
          "published_at": "2026-09-11T05:32:28.909844Z",
          "score": 4.761904761904762,
          "model": "gemini-3.8-flash",
          "runner": "benchmark_agent",
          "config": "google-gemini-3-8-flash-low",
          "verified_publish": true,
          "levels_completed": 1,
          "actions": 286,
          "resets": 3,
          "state": "GAME_OVER",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            115,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            16,
            270,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            29,
            54,
            85,
            300,
            208,
            113
          ]
        },
        {
          "session_id": "bc521569-90e5-44e4-9b36-02e07a819363",
          "game_id": "cn04",
          "game_version": "2fe56bfb",
          "published_at": "2026-09-10T20:36:36.291841Z",
          "score": 2.165906925236292,
          "model": "gemini-3.8-flash",
          "runner": "benchmark_agent",
          "config": "google-gemini-3-8-flash-medium",
          "verified_publish": true,
          "levels_completed": 1,
          "actions": 313,
          "resets": 4,
          "state": "GAME_OVER",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            45.48404542996214,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            43,
            270,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            29,
            54,
            85,
            300,
            208,
            113
          ]
        },
        {
          "session_id": "80b8124d-5354-4fc4-9201-93223f66ae95",
          "game_id": "cn04",
          "game_version": "2fe56bfb",
          "published_at": "2026-08-11T13:55:18.228416Z",
          "score": 4.761904761904762,
          "model": "grok-4.6",
          "runner": "benchmark_agent",
          "config": "xai-grok-4-6-xhigh",
          "verified_publish": true,
          "levels_completed": 1,
          "actions": 288,
          "resets": 2,
          "state": "GAME_OVER",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            115,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            18,
            270,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            29,
            54,
            85,
            300,
            208,
            113
          ]
        },
        {
          "session_id": "8b2c816a-e2c1-4c84-8ecf-15b6e4c335e2",
          "game_id": "cn04",
          "game_version": "2fe56bfb",
          "published_at": "2026-07-22T12:58:31.505156Z",
          "score": 47.61904761904761,
          "model": "claude-opus-5",
          "runner": "benchmark_agent",
          "config": "anthropic-claude-opus-5-high",
          "verified_publish": true,
          "levels_completed": 4,
          "actions": 519,
          "resets": 6,
          "state": "GAME_OVER",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            115,
            115,
            115,
            115,
            0,
            0
          ],
          "level_actions": [
            16,
            38,
            24,
            39,
            402,
            0
          ],
          "level_baseline_actions": [
            29,
            54,
            85,
            300,
            208,
            113
          ]
        },
        {
          "session_id": "09b4b506-009d-4e8b-80af-102f6777af17",
          "game_id": "cn04",
          "game_version": "2fe56bfb",
          "published_at": "2026-06-24T10:59:23.150575Z",
          "score": 0,
          "model": "gpt-5-6-luna",
          "runner": "benchmark_agent",
          "config": "openai-gpt-5-6-luna-low",
          "verified_publish": true,
          "levels_completed": 0,
          "actions": 145,
          "resets": 3,
          "state": "GAME_OVER",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            145,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            29,
            54,
            85,
            300,
            208,
            113
          ]
        },
        {
          "session_id": "30c56822-bd7a-48ac-937d-77973896c8ff",
          "game_id": "cn04",
          "game_version": "2fe56bfb",
          "published_at": "2026-06-24T10:40:56.749972Z",
          "score": 0,
          "model": "gpt-5-6-luna",
          "runner": "benchmark_agent",
          "config": "openai-gpt-5-6-luna-medium",
          "verified_publish": true,
          "levels_completed": 0,
          "actions": 145,
          "resets": 2,
          "state": "GAME_OVER",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            145,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            29,
            54,
            85,
            300,
            208,
            113
          ]
        },
        {
          "session_id": "9510e927-c6ba-4f18-a0d8-bb10d51628dd",
          "game_id": "cn04",
          "game_version": "2fe56bfb",
          "published_at": "2026-06-24T04:31:25.572587Z",
          "score": 0,
          "model": "gpt-5-6-luna",
          "runner": "benchmark_agent",
          "config": "openai-gpt-5-6-luna-high",
          "verified_publish": true,
          "levels_completed": 0,
          "actions": 145,
          "resets": 19,
          "state": "NOT_FINISHED",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            145,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            29,
            54,
            85,
            300,
            208,
            113
          ]
        },
        {
          "session_id": "7864cbad-3e35-4971-8eb1-9832c2657815",
          "game_id": "cn04",
          "game_version": "2fe56bfb",
          "published_at": "2026-06-24T02:52:39.065549Z",
          "score": 0,
          "model": "gpt-5-6-luna",
          "runner": "benchmark_agent",
          "config": "openai-gpt-5-6-luna-xhigh",
          "verified_publish": true,
          "levels_completed": 0,
          "actions": 145,
          "resets": 32,
          "state": "NOT_FINISHED",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            145,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            29,
            54,
            85,
            300,
            208,
            113
          ]
        },
        {
          "session_id": "fd77e211-49f7-4078-9d82-00504134ed34",
          "game_id": "cn04",
          "game_version": "2fe56bfb",
          "published_at": "2026-06-23T03:05:18.431812Z",
          "score": 0,
          "model": "gpt-5-6-luna",
          "runner": "benchmark_agent",
          "config": "openai-gpt-5-6-luna-max",
          "verified_publish": true,
          "levels_completed": 0,
          "actions": 145,
          "resets": 33,
          "state": "NOT_FINISHED",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            145,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            29,
            54,
            85,
            300,
            208,
            113
          ]
        },
        {
          "session_id": "1c6e9151-6dce-4d41-beca-4c9a078c388c",
          "game_id": "cn04",
          "game_version": "2fe56bfb",
          "published_at": "2026-06-20T12:06:29.83445Z",
          "score": 0,
          "model": "gpt-5-6-terra",
          "runner": "benchmark_agent",
          "config": "openai-gpt-5-6-terra-low",
          "verified_publish": true,
          "levels_completed": 0,
          "actions": 145,
          "resets": 1,
          "state": "GAME_OVER",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            145,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            29,
            54,
            85,
            300,
            208,
            113
          ]
        },
        {
          "session_id": "d3b35b30-430d-4a70-b51b-c0f5ca031db8",
          "game_id": "cn04",
          "game_version": "2fe56bfb",
          "published_at": "2026-06-20T09:39:04.959784Z",
          "score": 0,
          "model": "gpt-5-6-terra",
          "runner": "benchmark_agent",
          "config": "openai-gpt-5-6-terra-medium",
          "verified_publish": true,
          "levels_completed": 0,
          "actions": 145,
          "resets": 1,
          "state": "GAME_OVER",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            145,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            29,
            54,
            85,
            300,
            208,
            113
          ]
        },
        {
          "session_id": "5b5c211c-65ce-4715-bab0-db7a00a15421",
          "game_id": "cn04",
          "game_version": "2fe56bfb",
          "published_at": "2026-06-20T09:16:55.039267Z",
          "score": 0,
          "model": "gpt-5-6-terra",
          "runner": "benchmark_agent",
          "config": "openai-gpt-5-6-terra-high",
          "verified_publish": true,
          "levels_completed": 0,
          "actions": 145,
          "resets": 2,
          "state": "NOT_FINISHED",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            145,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            29,
            54,
            85,
            300,
            208,
            113
          ]
        },
        {
          "session_id": "4f1f89a8-69a7-4efb-b7ba-c62d1afeb483",
          "game_id": "cn04",
          "game_version": "2fe56bfb",
          "published_at": "2026-06-20T03:13:14.055086Z",
          "score": 0,
          "model": "gpt-5-6-terra",
          "runner": "benchmark_agent",
          "config": "openai-gpt-5-6-terra-xhigh",
          "verified_publish": true,
          "levels_completed": 0,
          "actions": 145,
          "resets": 3,
          "state": "GAME_OVER",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            145,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            29,
            54,
            85,
            300,
            208,
            113
          ]
        },
        {
          "session_id": "49a52862-c56b-4744-ab1e-a1538403f1c0",
          "game_id": "cn04",
          "game_version": "2fe56bfb",
          "published_at": "2026-06-19T04:05:28.958481Z",
          "score": 0.43454447751322745,
          "model": "gpt-5-6-terra",
          "runner": "benchmark_agent",
          "config": "openai-gpt-5-6-terra-max",
          "verified_publish": true,
          "levels_completed": 1,
          "actions": 366,
          "resets": 41,
          "state": "NOT_FINISHED",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            9.125434027777777,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            96,
            270,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            29,
            54,
            85,
            300,
            208,
            113
          ]
        },
        {
          "session_id": "24740104-9c45-46bd-8283-59f4d93a351e",
          "game_id": "cn04",
          "game_version": "2fe56bfb",
          "published_at": "2026-06-14T01:16:52.807724Z",
          "score": 31.74328299160415,
          "model": "gpt-5-6-sol",
          "runner": "benchmark_agent",
          "config": "openai-gpt-5-6-sol-max",
          "verified_publish": true,
          "levels_completed": 4,
          "actions": 911,
          "resets": 25,
          "state": "NOT_FINISHED",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            115,
            115,
            4.607310478522599,
            76.94675284702984,
            0,
            0
          ],
          "level_actions": [
            16,
            31,
            396,
            342,
            126,
            0
          ],
          "level_baseline_actions": [
            29,
            54,
            85,
            300,
            208,
            113
          ]
        },
        {
          "session_id": "87ea9d2e-acf1-4bd4-8f13-7232d0ad2bdf",
          "game_id": "cn04",
          "game_version": "2fe56bfb",
          "published_at": "2026-06-12T00:31:54.410987Z",
          "score": 4.761904761904762,
          "model": "gpt-5-6-sol",
          "runner": "benchmark_agent",
          "config": "openai-gpt-5-6-sol-xhigh",
          "verified_publish": true,
          "levels_completed": 1,
          "actions": 160,
          "resets": 5,
          "state": "NOT_FINISHED",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            115,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            22,
            138,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            29,
            54,
            85,
            300,
            208,
            113
          ]
        },
        {
          "session_id": "54faedc5-290d-4b51-bd5a-6ffd60dff526",
          "game_id": "cn04",
          "game_version": "2fe56bfb",
          "published_at": "2026-06-10T15:33:42.41298Z",
          "score": 4.761904761904762,
          "model": "gpt-5-6-sol",
          "runner": "benchmark_agent",
          "config": "openai-gpt-5-6-sol-high",
          "verified_publish": true,
          "levels_completed": 1,
          "actions": 294,
          "resets": 8,
          "state": "NOT_FINISHED",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            115,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            24,
            270,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            29,
            54,
            85,
            300,
            208,
            113
          ]
        },
        {
          "session_id": "ce135ddc-cfe0-4d9f-9563-088727ba0d89",
          "game_id": "cn04",
          "game_version": "2fe56bfb",
          "published_at": "2026-06-10T09:01:02.666182Z",
          "score": 0,
          "model": "gpt-5-6-sol",
          "runner": "benchmark_agent",
          "config": "openai-gpt-5-6-sol-medium",
          "verified_publish": true,
          "levels_completed": 0,
          "actions": 145,
          "resets": 1,
          "state": "GAME_OVER",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            145,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            29,
            54,
            85,
            300,
            208,
            113
          ]
        },
        {
          "session_id": "ee73a080-ff1f-4cfe-ae68-c323db55c654",
          "game_id": "cn04",
          "game_version": "2fe56bfb",
          "published_at": "2026-06-09T22:41:43.76247Z",
          "score": 0,
          "model": "gpt-5-6-sol",
          "runner": "benchmark_agent",
          "config": "openai-gpt-5-6-sol-low",
          "verified_publish": true,
          "levels_completed": 0,
          "actions": 145,
          "resets": 2,
          "state": "NOT_FINISHED",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            145,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            29,
            54,
            85,
            300,
            208,
            113
          ]
        },
        {
          "session_id": "e43ed5cd-7bf8-4107-8599-04d24a34a769",
          "game_id": "cn04",
          "game_version": "2fe56bfb",
          "published_at": "2026-05-29T20:15:35.876623Z",
          "score": 0.20143664326552507,
          "model": "claude-opus-4-8",
          "runner": "benchmark_agent",
          "config": "anthropic-opus-4-8-high",
          "verified_publish": true,
          "levels_completed": 1,
          "actions": 195,
          "resets": 5,
          "state": "GAME_OVER",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            4.230169508576027,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            141,
            54,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            29,
            54,
            85,
            300,
            208,
            113
          ]
        },
        {
          "session_id": "97f55fd4-3df9-4498-9373-3cc69a4695d3",
          "game_id": "cn04",
          "game_version": "2fe56bfb",
          "published_at": "2026-05-01T16:15:45.854728Z",
          "score": 0,
          "model": "gpt-5.5-2026-04-23",
          "runner": "benchmark_agent",
          "config": "openai-gpt-5-5-2026-04-23-high",
          "verified_publish": true,
          "levels_completed": 0,
          "actions": 145,
          "resets": 2,
          "state": "NOT_FINISHED",
          "completed": false,
          "number_of_levels": 6,
          "level_scores": [
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            145,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            29,
            54,
            85,
            300,
            208,
            113
          ]
        },
        {
          "session_id": "c99b68f5-0d61-46ff-a7b3-3491ef886f36",
          "game_id": "cn04",
          "game_version": "2fe56bfb",
          "published_at": "2026-05-01T05:08:52.890419Z",
          "score": 0,
          "model": "claude-opus-4-7",
          "runner": "benchmark_agent",
          "config": "anthropic-opus-4-7-high",
          "verified_publish": true,
          "levels_completed": 0,
          "actions": 145,
          "resets": 6,
          "state": "NOT_FINISHED",
          "completed": false,
          "number_of_levels": 6,
          "level_scores": [
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            145,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            29,
            54,
            85,
            300,
            208,
            113
          ]
        },
        {
          "session_id": "8f3fd2f0-c3de-4998-b02a-f737fc130f04",
          "game_id": "cn04",
          "game_version": "2fe56bfb",
          "published_at": "2026-04-14T17:44:30.842158Z",
          "score": 0,
          "model": "grok-4.20-beta-0309-reasoning",
          "runner": "benchmark_agent",
          "config": "xai-grok-4-20-beta-0309-reasoning",
          "verified_publish": true,
          "levels_completed": 0,
          "actions": 145,
          "resets": 1,
          "state": "GAME_OVER",
          "completed": false,
          "number_of_levels": 6,
          "level_scores": [
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            145,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            29,
            54,
            85,
            300,
            208,
            113
          ]
        },
        {
          "session_id": "e9db3254-ab1c-42ae-9ae4-408549b57c0f",
          "game_id": "cn04",
          "game_version": "2fe56bfb",
          "published_at": "2026-04-14T17:44:30.55108Z",
          "score": 0,
          "model": "gpt-5.4-2026-03-05",
          "runner": "benchmark_agent",
          "config": "openai-gpt-5-4-2026-03-05-high",
          "verified_publish": true,
          "levels_completed": 0,
          "actions": 145,
          "resets": 2,
          "state": "GAME_OVER",
          "completed": false,
          "number_of_levels": 6,
          "level_scores": [
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            145,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            29,
            54,
            85,
            300,
            208,
            113
          ]
        },
        {
          "session_id": "73e6b589-f10b-4b4b-b9f4-28dbe0012b3c",
          "game_id": "cn04",
          "game_version": "2fe56bfb",
          "published_at": "2026-04-14T17:44:29.882789Z",
          "score": 0,
          "model": "gemini-3.1-pro-preview",
          "runner": "benchmark_agent",
          "config": "google-gemini-3-1-pro-preview",
          "verified_publish": true,
          "levels_completed": 0,
          "actions": 53,
          "resets": 2,
          "state": "NOT_FINISHED",
          "completed": false,
          "number_of_levels": 6,
          "level_scores": [
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            53,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            29,
            54,
            85,
            300,
            208,
            113
          ]
        },
        {
          "session_id": "8ee029d9-d607-4cf0-8f19-0bff3bc537af",
          "game_id": "cn04",
          "game_version": "2fe56bfb",
          "published_at": "2026-04-14T17:44:29.419756Z",
          "score": 0,
          "model": "claude-opus-4-6",
          "runner": "benchmark_agent",
          "config": "anthropic-opus-4-6-max-effort",
          "verified_publish": true,
          "levels_completed": 0,
          "actions": 145,
          "resets": 5,
          "state": "NOT_FINISHED",
          "completed": false,
          "number_of_levels": 6,
          "level_scores": [
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            145,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            29,
            54,
            85,
            300,
            208,
            113
          ]
        }
      ]
    },
    {
      "game": "m0r0",
      "source": "https://arcprize.org/tasks/m0r0",
      "runs": [
        {
          "session_id": "474503f7-2af7-416a-bdd5-236e4879f5ac",
          "game_id": "m0r0",
          "game_version": "492f87ba",
          "published_at": "2026-09-25T13:35:13.761953Z",
          "score": 4.185267857142857,
          "model": "gpt-6-sol",
          "runner": "benchmark_agent",
          "config": "openai-gpt-6-sol-max",
          "verified_publish": true,
          "levels_completed": 1,
          "actions": 587,
          "resets": 3,
          "state": "GAME_OVER",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            87.890625,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            32,
            555,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            30,
            111,
            203,
            26,
            500,
            237
          ]
        },
        {
          "session_id": "004b4a36-34bc-4b70-a795-4e8f9ce5a5ed",
          "game_id": "m0r0",
          "game_version": "492f87ba",
          "published_at": "2026-09-25T07:56:45.576207Z",
          "score": 14.285714285714285,
          "model": "gpt-6-sol",
          "runner": "benchmark_agent",
          "config": "openai-gpt-6-sol-low-provider-adapter",
          "verified_publish": true,
          "levels_completed": 2,
          "actions": 1145,
          "resets": 8,
          "state": "GAME_OVER",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            82.64462809917354,
            115,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            33,
            97,
            1015,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            30,
            111,
            203,
            26,
            500,
            237
          ]
        },
        {
          "session_id": "43575160-1b49-4523-a7b3-8bff36adea8f",
          "game_id": "m0r0",
          "game_version": "492f87ba",
          "published_at": "2026-09-25T07:50:30.577042Z",
          "score": 12.369146005509641,
          "model": "gpt-6-sol",
          "runner": "benchmark_agent",
          "config": "openai-gpt-6-sol-medium-provider-adapter",
          "verified_publish": true,
          "levels_completed": 2,
          "actions": 1151,
          "resets": 9,
          "state": "GAME_OVER",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            29.752066115702476,
            115,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            55,
            81,
            1015,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            30,
            111,
            203,
            26,
            500,
            237
          ]
        },
        {
          "session_id": "a60911fd-9fc6-405f-b37f-114d6c7e9d93",
          "game_id": "m0r0",
          "game_version": "492f87ba",
          "published_at": "2026-09-25T07:45:25.082673Z",
          "score": 4.4596402556860415,
          "model": "gpt-6-sol",
          "runner": "benchmark_agent",
          "config": "openai-gpt-6-sol-xhigh",
          "verified_publish": true,
          "levels_completed": 1,
          "actions": 586,
          "resets": 3,
          "state": "GAME_OVER",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            93.65244536940688,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            31,
            555,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            30,
            111,
            203,
            26,
            500,
            237
          ]
        },
        {
          "session_id": "5d9c8117-71d2-4ca0-b9c3-21db7dbd61fa",
          "game_id": "m0r0",
          "game_version": "492f87ba",
          "published_at": "2026-09-25T07:15:35.834645Z",
          "score": 14.285714285714285,
          "model": "gpt-6-sol",
          "runner": "benchmark_agent",
          "config": "openai-gpt-6-sol-high-provider-adapter",
          "verified_publish": true,
          "levels_completed": 2,
          "actions": 1145,
          "resets": 8,
          "state": "GAME_OVER",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            73.46938775510203,
            115,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            35,
            95,
            1015,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            30,
            111,
            203,
            26,
            500,
            237
          ]
        },
        {
          "session_id": "325e8446-609c-4645-9fe2-a855f5862861",
          "game_id": "m0r0",
          "game_version": "492f87ba",
          "published_at": "2026-09-25T07:04:35.311026Z",
          "score": 0.3077933270406697,
          "model": "gpt-6-sol",
          "runner": "benchmark_agent",
          "config": "openai-gpt-6-sol-low",
          "verified_publish": true,
          "levels_completed": 1,
          "actions": 673,
          "resets": 5,
          "state": "GAME_OVER",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            6.463659867854064,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            118,
            555,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            30,
            111,
            203,
            26,
            500,
            237
          ]
        },
        {
          "session_id": "2e466d89-ab79-44f0-b492-8fde27793a23",
          "game_id": "m0r0",
          "game_version": "492f87ba",
          "published_at": "2026-09-25T06:55:25.936283Z",
          "score": 2.3178552113111337,
          "model": "gpt-6-sol",
          "runner": "benchmark_agent",
          "config": "openai-gpt-6-sol-none",
          "verified_publish": true,
          "levels_completed": 1,
          "actions": 598,
          "resets": 5,
          "state": "GAME_OVER",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            48.67495943753381,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            43,
            555,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            30,
            111,
            203,
            26,
            500,
            237
          ]
        },
        {
          "session_id": "0a70c363-0d36-4922-b322-8d069acfa883",
          "game_id": "m0r0",
          "game_version": "492f87ba",
          "published_at": "2026-09-25T06:46:19.498179Z",
          "score": 3.4985422740524776,
          "model": "gpt-6-sol",
          "runner": "benchmark_agent",
          "config": "openai-gpt-6-sol-medium",
          "verified_publish": true,
          "levels_completed": 1,
          "actions": 590,
          "resets": 3,
          "state": "GAME_OVER",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            73.46938775510203,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            35,
            555,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            30,
            111,
            203,
            26,
            500,
            237
          ]
        },
        {
          "session_id": "35df4b2c-a701-4083-8843-4dae3b0bad35",
          "game_id": "m0r0",
          "game_version": "492f87ba",
          "published_at": "2026-09-25T06:46:09.705591Z",
          "score": 1.2739935450993716,
          "model": "gpt-6-sol",
          "runner": "benchmark_agent",
          "config": "openai-gpt-6-sol-high",
          "verified_publish": true,
          "levels_completed": 1,
          "actions": 613,
          "resets": 3,
          "state": "GAME_OVER",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            26.753864447086805,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            58,
            555,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            30,
            111,
            203,
            26,
            500,
            237
          ]
        },
        {
          "session_id": "52110493-9d0d-4d35-a4ef-04a8daa0f52f",
          "game_id": "m0r0",
          "game_version": "492f87ba",
          "published_at": "2026-09-25T06:46:01.626626Z",
          "score": 2.3178552113111337,
          "model": "gpt-6-sol",
          "runner": "benchmark_agent",
          "config": "openai-gpt-6-sol-none-provider-adapter",
          "verified_publish": true,
          "levels_completed": 1,
          "actions": 186,
          "resets": 2,
          "state": "NOT_FINISHED",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            48.67495943753381,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            43,
            143,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            30,
            111,
            203,
            26,
            500,
            237
          ]
        },
        {
          "session_id": "971c5ba3-aac0-48a7-b03d-f50976c65dd0",
          "game_id": "m0r0",
          "game_version": "492f87ba",
          "published_at": "2026-09-25T06:30:02.697421Z",
          "score": 14.082924623465162,
          "model": "gpt-6-sol",
          "runner": "benchmark_agent",
          "config": "openai-gpt-6-sol-xhigh-provider-adapter",
          "verified_publish": true,
          "levels_completed": 2,
          "actions": 1088,
          "resets": 9,
          "state": "GAME_OVER",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            65.74141709276844,
            115,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            37,
            36,
            1015,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            30,
            111,
            203,
            26,
            500,
            237
          ]
        },
        {
          "session_id": "a83c70e8-399c-4ddb-9db5-19be40b67c22",
          "game_id": "m0r0",
          "game_version": "492f87ba",
          "published_at": "2026-09-21T11:46:18.646162Z",
          "score": 28.57142857142857,
          "model": "gemini-3.8-flash",
          "runner": "benchmark_agent",
          "config": "google-gemini-3-8-flash-low-provider-adapter",
          "verified_publish": true,
          "levels_completed": 3,
          "actions": 332,
          "resets": 0,
          "state": "NOT_FINISHED",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            115,
            115,
            115,
            0,
            0,
            0
          ],
          "level_actions": [
            26,
            41,
            135,
            130,
            0,
            0
          ],
          "level_baseline_actions": [
            30,
            111,
            203,
            26,
            500,
            237
          ]
        },
        {
          "session_id": "596f83e4-12d3-4d0d-8247-177bd3ecd163",
          "game_id": "m0r0",
          "game_version": "492f87ba",
          "published_at": "2026-09-20T08:49:26.091696Z",
          "score": 59.75214283164076,
          "model": "gemini-3.8-flash",
          "runner": "benchmark_agent",
          "config": "google-gemini-3-8-flash-medium-provider-adapter",
          "verified_publish": true,
          "levels_completed": 6,
          "actions": 1011,
          "resets": 6,
          "state": "WIN",
          "completed": true,
          "number_of_levels": 0,
          "level_scores": [
            115,
            115,
            23.36111111111111,
            5.29407157960686,
            115,
            40.58922996878251
          ],
          "level_actions": [
            20,
            23,
            420,
            113,
            63,
            372
          ],
          "level_baseline_actions": [
            30,
            111,
            203,
            26,
            500,
            237
          ]
        },
        {
          "session_id": "d1074fdc-317d-458d-a410-ba681d8b5030",
          "game_id": "m0r0",
          "game_version": "492f87ba",
          "published_at": "2026-09-18T02:50:21.891692Z",
          "score": 59.91604621057801,
          "model": "gemini-3.8-flash",
          "runner": "benchmark_agent",
          "config": "google-gemini-3-8-flash-high-provider-adapter",
          "verified_publish": true,
          "levels_completed": 5,
          "actions": 1346,
          "resets": 19,
          "state": "NOT_FINISHED",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            115,
            115,
            59.126779155188245,
            40.214158239143366,
            115,
            0
          ],
          "level_actions": [
            20,
            23,
            264,
            41,
            61,
            937
          ],
          "level_baseline_actions": [
            30,
            111,
            203,
            26,
            500,
            237
          ]
        },
        {
          "session_id": "42e8ee9a-1ed2-4182-bc70-0927553f5268",
          "game_id": "m0r0",
          "game_version": "492f87ba",
          "published_at": "2026-09-16T19:39:14.865739Z",
          "score": 0,
          "model": "gpt-6-luna",
          "runner": "benchmark_agent",
          "config": "openai-gpt-6-luna-medium",
          "verified_publish": true,
          "levels_completed": 0,
          "actions": 150,
          "resets": 3,
          "state": "NOT_FINISHED",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            150,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            30,
            111,
            203,
            26,
            500,
            237
          ]
        },
        {
          "session_id": "c0da8873-06b8-475e-a79f-7338cb932c15",
          "game_id": "m0r0",
          "game_version": "492f87ba",
          "published_at": "2026-09-16T19:33:20.81782Z",
          "score": 0,
          "model": "gpt-6-luna",
          "runner": "benchmark_agent",
          "config": "openai-gpt-6-luna-medium-provider-adapter",
          "verified_publish": true,
          "levels_completed": 0,
          "actions": 150,
          "resets": 4,
          "state": "NOT_FINISHED",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            150,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            30,
            111,
            203,
            26,
            500,
            237
          ]
        },
        {
          "session_id": "2cb597de-8503-4453-b329-10dbc105d72c",
          "game_id": "m0r0",
          "game_version": "492f87ba",
          "published_at": "2026-09-16T19:26:46.999999Z",
          "score": 0,
          "model": "gpt-6-luna",
          "runner": "benchmark_agent",
          "config": "openai-gpt-6-luna-low",
          "verified_publish": true,
          "levels_completed": 0,
          "actions": 150,
          "resets": 0,
          "state": "NOT_FINISHED",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            150,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            30,
            111,
            203,
            26,
            500,
            237
          ]
        },
        {
          "session_id": "920c19f1-d29d-4212-9a1b-bf39baa7557e",
          "game_id": "m0r0",
          "game_version": "492f87ba",
          "published_at": "2026-09-16T19:25:59.324574Z",
          "score": 0,
          "model": "gpt-6-luna",
          "runner": "benchmark_agent",
          "config": "openai-gpt-6-luna-low-provider-adapter",
          "verified_publish": true,
          "levels_completed": 0,
          "actions": 150,
          "resets": 1,
          "state": "NOT_FINISHED",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            150,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            30,
            111,
            203,
            26,
            500,
            237
          ]
        },
        {
          "session_id": "af6bbf5e-c53c-479b-8974-b9f1af637b0f",
          "game_id": "m0r0",
          "game_version": "492f87ba",
          "published_at": "2026-09-16T19:23:58.062374Z",
          "score": 0,
          "model": "gpt-6-luna",
          "runner": "benchmark_agent",
          "config": "openai-gpt-6-luna-none",
          "verified_publish": true,
          "levels_completed": 0,
          "actions": 150,
          "resets": 0,
          "state": "NOT_FINISHED",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            150,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            30,
            111,
            203,
            26,
            500,
            237
          ]
        },
        {
          "session_id": "f620678b-81cb-4a68-a08d-8bebb0e5421c",
          "game_id": "m0r0",
          "game_version": "492f87ba",
          "published_at": "2026-09-16T16:12:45.778133Z",
          "score": 0,
          "model": "gpt-6-luna",
          "runner": "benchmark_agent",
          "config": "openai-gpt-6-luna-high",
          "verified_publish": true,
          "levels_completed": 0,
          "actions": 150,
          "resets": 3,
          "state": "NOT_FINISHED",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            150,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            30,
            111,
            203,
            26,
            500,
            237
          ]
        },
        {
          "session_id": "fd7611f0-c617-45d4-af78-d9a50ce443e4",
          "game_id": "m0r0",
          "game_version": "492f87ba",
          "published_at": "2026-09-16T09:47:17.450002Z",
          "score": 0,
          "model": "gpt-6-luna",
          "runner": "benchmark_agent",
          "config": "openai-gpt-6-luna-max",
          "verified_publish": true,
          "levels_completed": 0,
          "actions": 150,
          "resets": 9,
          "state": "NOT_FINISHED",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            150,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            30,
            111,
            203,
            26,
            500,
            237
          ]
        },
        {
          "session_id": "97265932-dc4b-498b-8faa-c45807d9bf96",
          "game_id": "m0r0",
          "game_version": "492f87ba",
          "published_at": "2026-09-16T09:15:04.54141Z",
          "score": 0,
          "model": "gpt-6-luna",
          "runner": "benchmark_agent",
          "config": "openai-gpt-6-luna-none-provider-adapter",
          "verified_publish": true,
          "levels_completed": 0,
          "actions": 150,
          "resets": 3,
          "state": "NOT_FINISHED",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            150,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            30,
            111,
            203,
            26,
            500,
            237
          ]
        },
        {
          "session_id": "69e99852-3e30-4cbd-8fd8-a8a02eab5d4e",
          "game_id": "m0r0",
          "game_version": "492f87ba",
          "published_at": "2026-09-16T09:09:36.459332Z",
          "score": 0.5534238488783944,
          "model": "gpt-6-luna",
          "runner": "benchmark_agent",
          "config": "openai-gpt-6-luna-xhigh-provider-adapter",
          "verified_publish": true,
          "levels_completed": 1,
          "actions": 446,
          "resets": 5,
          "state": "GAME_OVER",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            11.62190082644628,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            88,
            358,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            30,
            111,
            203,
            26,
            500,
            237
          ]
        },
        {
          "session_id": "f4f87f1c-77f6-4b1a-85e3-c15778cac633",
          "game_id": "m0r0",
          "game_version": "492f87ba",
          "published_at": "2026-09-16T08:54:13.416508Z",
          "score": 0,
          "model": "gpt-6-luna",
          "runner": "benchmark_agent",
          "config": "openai-gpt-6-luna-high-provider-adapter",
          "verified_publish": true,
          "levels_completed": 0,
          "actions": 150,
          "resets": 2,
          "state": "NOT_FINISHED",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            150,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            30,
            111,
            203,
            26,
            500,
            237
          ]
        },
        {
          "session_id": "2bc7429a-ed03-4604-9aaf-36839dbe2c1a",
          "game_id": "m0r0",
          "game_version": "492f87ba",
          "published_at": "2026-09-16T08:28:27.228758Z",
          "score": 0,
          "model": "gpt-6-luna",
          "runner": "benchmark_agent",
          "config": "openai-gpt-6-luna-xhigh",
          "verified_publish": true,
          "levels_completed": 0,
          "actions": 150,
          "resets": 4,
          "state": "NOT_FINISHED",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            150,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            30,
            111,
            203,
            26,
            500,
            237
          ]
        },
        {
          "session_id": "421c5270-c684-4a10-95af-c43d68cc4a86",
          "game_id": "m0r0",
          "game_version": "492f87ba",
          "published_at": "2026-09-16T08:11:40.681625Z",
          "score": 14.285714285714285,
          "model": "gpt-6-1-sol",
          "runner": "benchmark_agent",
          "config": "openai-gpt-6-1-sol-low",
          "verified_publish": true,
          "levels_completed": 2,
          "actions": 1060,
          "resets": 6,
          "state": "GAME_OVER",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            115,
            115,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            20,
            25,
            1015,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            30,
            111,
            203,
            26,
            500,
            237
          ]
        },
        {
          "session_id": "e973763b-e4a5-4179-86c7-8bf8cb0fd57e",
          "game_id": "m0r0",
          "game_version": "492f87ba",
          "published_at": "2026-09-16T07:07:18.321712Z",
          "score": 14.285714285714285,
          "model": "gpt-6-1-sol",
          "runner": "benchmark_agent",
          "config": "openai-gpt-6-1-sol-medium",
          "verified_publish": true,
          "levels_completed": 2,
          "actions": 1054,
          "resets": 6,
          "state": "GAME_OVER",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            115,
            115,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            16,
            23,
            1015,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            30,
            111,
            203,
            26,
            500,
            237
          ]
        },
        {
          "session_id": "51faa25c-1cbf-40dd-8c1d-236de657dc66",
          "game_id": "m0r0",
          "game_version": "492f87ba",
          "published_at": "2026-09-16T06:56:58.956419Z",
          "score": 14.285714285714285,
          "model": "gpt-6-1-sol",
          "runner": "benchmark_agent",
          "config": "openai-gpt-6-1-sol-high",
          "verified_publish": true,
          "levels_completed": 2,
          "actions": 1079,
          "resets": 7,
          "state": "GAME_OVER",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            115,
            115,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            18,
            46,
            1015,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            30,
            111,
            203,
            26,
            500,
            237
          ]
        },
        {
          "session_id": "ca789fe3-1418-4517-aa39-4df3fc7712ce",
          "game_id": "m0r0",
          "game_version": "492f87ba",
          "published_at": "2026-09-16T06:19:39.551696Z",
          "score": 71.42857142857143,
          "model": "gpt-6-1-sol",
          "runner": "benchmark_agent",
          "config": "openai-gpt-6-1-sol-max",
          "verified_publish": true,
          "levels_completed": 5,
          "actions": 275,
          "resets": 3,
          "state": "NOT_FINISHED",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            115,
            115,
            115,
            115,
            115,
            0
          ],
          "level_actions": [
            15,
            23,
            100,
            24,
            45,
            68
          ],
          "level_baseline_actions": [
            30,
            111,
            203,
            26,
            500,
            237
          ]
        },
        {
          "session_id": "66f8c04c-586a-4c0a-bc17-b4904b61b35f",
          "game_id": "m0r0",
          "game_version": "492f87ba",
          "published_at": "2026-09-16T05:10:22.246178Z",
          "score": 28.57142857142857,
          "model": "gpt-6-1-sol",
          "runner": "benchmark_agent",
          "config": "openai-gpt-6-1-sol-xhigh",
          "verified_publish": true,
          "levels_completed": 3,
          "actions": 367,
          "resets": 1,
          "state": "NOT_FINISHED",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            115,
            115,
            105.114274053668,
            0,
            0,
            0
          ],
          "level_actions": [
            16,
            23,
            198,
            130,
            0,
            0
          ],
          "level_baseline_actions": [
            30,
            111,
            203,
            26,
            500,
            237
          ]
        },
        {
          "session_id": "7716810a-7686-45bd-8b78-d3cc98c5c0e9",
          "game_id": "m0r0",
          "game_version": "492f87ba",
          "published_at": "2026-09-16T05:06:07.317383Z",
          "score": 100,
          "model": "gpt-6-1-sol",
          "runner": "benchmark_agent",
          "config": "openai-gpt-6-1-sol-high-provider-adapter",
          "verified_publish": true,
          "levels_completed": 6,
          "actions": 256,
          "resets": 0,
          "state": "WIN",
          "completed": true,
          "number_of_levels": 0,
          "level_scores": [
            115,
            115,
            115,
            115,
            115,
            115
          ],
          "level_actions": [
            15,
            23,
            85,
            23,
            58,
            52
          ],
          "level_baseline_actions": [
            30,
            111,
            203,
            26,
            500,
            237
          ]
        },
        {
          "session_id": "331f003d-f721-44b5-be50-c1dfd1c9f697",
          "game_id": "m0r0",
          "game_version": "492f87ba",
          "published_at": "2026-09-16T03:51:23.163584Z",
          "score": 100,
          "model": "gpt-6-1-sol",
          "runner": "benchmark_agent",
          "config": "openai-gpt-6-1-sol-xhigh-provider-adapter",
          "verified_publish": true,
          "levels_completed": 6,
          "actions": 253,
          "resets": 0,
          "state": "WIN",
          "completed": true,
          "number_of_levels": 0,
          "level_scores": [
            115,
            115,
            115,
            115,
            115,
            115
          ],
          "level_actions": [
            15,
            23,
            77,
            21,
            66,
            51
          ],
          "level_baseline_actions": [
            30,
            111,
            203,
            26,
            500,
            237
          ]
        },
        {
          "session_id": "628240a2-adc6-4611-a359-7e1cfcfced06",
          "game_id": "m0r0",
          "game_version": "492f87ba",
          "published_at": "2026-09-16T03:39:56.124611Z",
          "score": 100,
          "model": "gpt-6-1-sol",
          "runner": "benchmark_agent",
          "config": "openai-gpt-6-1-sol-max-provider-adapter",
          "verified_publish": true,
          "levels_completed": 6,
          "actions": 248,
          "resets": 0,
          "state": "WIN",
          "completed": true,
          "number_of_levels": 0,
          "level_scores": [
            115,
            115,
            115,
            115,
            115,
            115
          ],
          "level_actions": [
            15,
            23,
            65,
            15,
            53,
            77
          ],
          "level_baseline_actions": [
            30,
            111,
            203,
            26,
            500,
            237
          ]
        },
        {
          "session_id": "49406b29-4a21-47e6-adfb-f9b27e61db91",
          "game_id": "m0r0",
          "game_version": "492f87ba",
          "published_at": "2026-09-16T02:49:10.499843Z",
          "score": 0,
          "model": "gpt-6-luna",
          "runner": "benchmark_agent",
          "config": "openai-gpt-6-luna-max-provider-adapter",
          "verified_publish": true,
          "levels_completed": 0,
          "actions": 150,
          "resets": 0,
          "state": "NOT_FINISHED",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            150,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            30,
            111,
            203,
            26,
            500,
            237
          ]
        },
        {
          "session_id": "447466a1-c140-490d-84fc-6fc2fe489124",
          "game_id": "m0r0",
          "game_version": "492f87ba",
          "published_at": "2026-09-12T13:34:58.249401Z",
          "score": 7.733478432924417,
          "model": "gemini-3.8-flash",
          "runner": "benchmark_agent",
          "config": "google-gemini-3-8-flash-high",
          "verified_publish": true,
          "levels_completed": 2,
          "actions": 1268,
          "resets": 8,
          "state": "GAME_OVER",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            115,
            23.701523545706372,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            25,
            228,
            1015,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            30,
            111,
            203,
            26,
            500,
            237
          ]
        },
        {
          "session_id": "0a8b355e-dc26-49d4-acfb-138de630d040",
          "game_id": "m0r0",
          "game_version": "492f87ba",
          "published_at": "2026-09-12T08:59:41.559962Z",
          "score": 14.285714285714285,
          "model": "gemini-3.8-flash",
          "runner": "benchmark_agent",
          "config": "google-gemini-3-8-flash-low",
          "verified_publish": true,
          "levels_completed": 2,
          "actions": 1126,
          "resets": 7,
          "state": "GAME_OVER",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            100,
            115,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            30,
            81,
            1015,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            30,
            111,
            203,
            26,
            500,
            237
          ]
        },
        {
          "session_id": "7bf6fe4e-8f9a-4286-8fe3-b4d2f7111c89",
          "game_id": "m0r0",
          "game_version": "492f87ba",
          "published_at": "2026-09-11T04:04:31.027848Z",
          "score": 7.511498641579483,
          "model": "gemini-3.8-flash",
          "runner": "benchmark_agent",
          "config": "google-gemini-3-8-flash-medium",
          "verified_publish": true,
          "levels_completed": 2,
          "actions": 1221,
          "resets": 10,
          "state": "GAME_OVER",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            73.46938775510203,
            42.136041859033554,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            35,
            171,
            1015,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            30,
            111,
            203,
            26,
            500,
            237
          ]
        },
        {
          "session_id": "af6d0952-79f9-4607-814f-e57d36e79812",
          "game_id": "m0r0",
          "game_version": "492f87ba",
          "published_at": "2026-08-11T22:22:43.318967Z",
          "score": 0.2386786748559972,
          "model": "grok-4.6",
          "runner": "benchmark_agent",
          "config": "xai-grok-4-6-xhigh",
          "verified_publish": true,
          "levels_completed": 1,
          "actions": 689,
          "resets": 4,
          "state": "GAME_OVER",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            5.012252171975941,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            134,
            555,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            30,
            111,
            203,
            26,
            500,
            237
          ]
        },
        {
          "session_id": "e62c81b1-42d1-49ff-aa34-faede02c2e42",
          "game_id": "m0r0",
          "game_version": "492f87ba",
          "published_at": "2026-07-21T12:15:34.751174Z",
          "score": 28.57142857142857,
          "model": "claude-opus-5",
          "runner": "benchmark_agent",
          "config": "anthropic-claude-opus-5-high",
          "verified_publish": true,
          "levels_completed": 3,
          "actions": 270,
          "resets": 0,
          "state": "NOT_FINISHED",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            115,
            115,
            115,
            0,
            0,
            0
          ],
          "level_actions": [
            19,
            23,
            98,
            130,
            0,
            0
          ],
          "level_baseline_actions": [
            30,
            111,
            203,
            26,
            500,
            237
          ]
        },
        {
          "session_id": "3e77bc91-b4b0-481d-aaee-e320f8e08f6b",
          "game_id": "m0r0",
          "game_version": "492f87ba",
          "published_at": "2026-06-24T11:00:21.682262Z",
          "score": 0,
          "model": "gpt-5-6-luna",
          "runner": "benchmark_agent",
          "config": "openai-gpt-5-6-luna-low",
          "verified_publish": true,
          "levels_completed": 0,
          "actions": 150,
          "resets": 0,
          "state": "NOT_FINISHED",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            150,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            30,
            111,
            203,
            26,
            500,
            237
          ]
        },
        {
          "session_id": "55a1985f-c70c-4c8e-88e1-6e435e027166",
          "game_id": "m0r0",
          "game_version": "492f87ba",
          "published_at": "2026-06-24T10:42:25.962371Z",
          "score": 0,
          "model": "gpt-5-6-luna",
          "runner": "benchmark_agent",
          "config": "openai-gpt-5-6-luna-medium",
          "verified_publish": true,
          "levels_completed": 0,
          "actions": 150,
          "resets": 1,
          "state": "NOT_FINISHED",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            150,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            30,
            111,
            203,
            26,
            500,
            237
          ]
        },
        {
          "session_id": "47d20663-eff2-483a-903f-22afe80ee890",
          "game_id": "m0r0",
          "game_version": "492f87ba",
          "published_at": "2026-06-24T04:31:50.90274Z",
          "score": 0,
          "model": "gpt-5-6-luna",
          "runner": "benchmark_agent",
          "config": "openai-gpt-5-6-luna-high",
          "verified_publish": true,
          "levels_completed": 0,
          "actions": 150,
          "resets": 9,
          "state": "NOT_FINISHED",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            150,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            30,
            111,
            203,
            26,
            500,
            237
          ]
        },
        {
          "session_id": "0b3e0a79-c01a-4a52-8e2f-880d915b2cef",
          "game_id": "m0r0",
          "game_version": "492f87ba",
          "published_at": "2026-06-24T03:08:38.900037Z",
          "score": 0,
          "model": "gpt-5-6-luna",
          "runner": "benchmark_agent",
          "config": "openai-gpt-5-6-luna-xhigh",
          "verified_publish": true,
          "levels_completed": 0,
          "actions": 150,
          "resets": 10,
          "state": "NOT_FINISHED",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            150,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            30,
            111,
            203,
            26,
            500,
            237
          ]
        },
        {
          "session_id": "b39bb22d-88d2-4bb2-ae33-1bf92f42b9d5",
          "game_id": "m0r0",
          "game_version": "492f87ba",
          "published_at": "2026-06-23T02:50:18.208087Z",
          "score": 0,
          "model": "gpt-5-6-luna",
          "runner": "benchmark_agent",
          "config": "openai-gpt-5-6-luna-max",
          "verified_publish": true,
          "levels_completed": 0,
          "actions": 150,
          "resets": 23,
          "state": "NOT_FINISHED",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            150,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            30,
            111,
            203,
            26,
            500,
            237
          ]
        },
        {
          "session_id": "df08dc19-7e94-4850-8540-e9131725ebd4",
          "game_id": "m0r0",
          "game_version": "492f87ba",
          "published_at": "2026-06-20T12:06:10.801233Z",
          "score": 0,
          "model": "gpt-5-6-terra",
          "runner": "benchmark_agent",
          "config": "openai-gpt-5-6-terra-low",
          "verified_publish": true,
          "levels_completed": 0,
          "actions": 150,
          "resets": 1,
          "state": "NOT_FINISHED",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            150,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            30,
            111,
            203,
            26,
            500,
            237
          ]
        },
        {
          "session_id": "d8761d10-5028-4c43-92d8-662a48e7b4a8",
          "game_id": "m0r0",
          "game_version": "492f87ba",
          "published_at": "2026-06-20T11:01:10.375551Z",
          "score": 0.5291005291005291,
          "model": "gpt-5-6-terra",
          "runner": "benchmark_agent",
          "config": "openai-gpt-5-6-terra-high",
          "verified_publish": true,
          "levels_completed": 1,
          "actions": 645,
          "resets": 5,
          "state": "GAME_OVER",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            11.11111111111111,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            90,
            555,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            30,
            111,
            203,
            26,
            500,
            237
          ]
        },
        {
          "session_id": "ab1c41b5-a765-4a48-9eb1-f1dac69fd2d1",
          "game_id": "m0r0",
          "game_version": "492f87ba",
          "published_at": "2026-06-20T09:38:49.872689Z",
          "score": 0,
          "model": "gpt-5-6-terra",
          "runner": "benchmark_agent",
          "config": "openai-gpt-5-6-terra-medium",
          "verified_publish": true,
          "levels_completed": 0,
          "actions": 150,
          "resets": 0,
          "state": "NOT_FINISHED",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            150,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            30,
            111,
            203,
            26,
            500,
            237
          ]
        },
        {
          "session_id": "ab122113-c06f-4742-a629-789828d210a9",
          "game_id": "m0r0",
          "game_version": "492f87ba",
          "published_at": "2026-06-20T04:31:14.408372Z",
          "score": 0,
          "model": "gpt-5-6-terra",
          "runner": "benchmark_agent",
          "config": "openai-gpt-5-6-terra-xhigh",
          "verified_publish": true,
          "levels_completed": 0,
          "actions": 150,
          "resets": 7,
          "state": "NOT_FINISHED",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            150,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            30,
            111,
            203,
            26,
            500,
            237
          ]
        },
        {
          "session_id": "b5ba31a1-52b5-48e9-ab49-bc3459cca17e",
          "game_id": "m0r0",
          "game_version": "492f87ba",
          "published_at": "2026-06-19T08:52:57.369395Z",
          "score": 0.8042248612712113,
          "model": "gpt-5-6-terra",
          "runner": "benchmark_agent",
          "config": "openai-gpt-5-6-terra-max",
          "verified_publish": true,
          "levels_completed": 1,
          "actions": 628,
          "resets": 8,
          "state": "GAME_OVER",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            16.888722086695438,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            73,
            555,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            30,
            111,
            203,
            26,
            500,
            237
          ]
        },
        {
          "session_id": "68a428df-672d-4995-bb2b-20ffd18a4368",
          "game_id": "m0r0",
          "game_version": "492f87ba",
          "published_at": "2026-06-13T14:40:50.162173Z",
          "score": 1.2739935450993716,
          "model": "gpt-5-6-sol",
          "runner": "benchmark_agent",
          "config": "openai-gpt-5-6-sol-max",
          "verified_publish": true,
          "levels_completed": 1,
          "actions": 613,
          "resets": 3,
          "state": "GAME_OVER",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            26.753864447086805,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            58,
            555,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            30,
            111,
            203,
            26,
            500,
            237
          ]
        },
        {
          "session_id": "6d1afe52-4af6-4519-a2a7-70f125b0101f",
          "game_id": "m0r0",
          "game_version": "492f87ba",
          "published_at": "2026-06-12T00:32:13.770883Z",
          "score": 4.761904761904762,
          "model": "gpt-5-6-sol",
          "runner": "benchmark_agent",
          "config": "openai-gpt-5-6-sol-xhigh",
          "verified_publish": true,
          "levels_completed": 1,
          "actions": 189,
          "resets": 3,
          "state": "GAME_OVER",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            115,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            27,
            162,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            30,
            111,
            203,
            26,
            500,
            237
          ]
        },
        {
          "session_id": "ea3d9d86-86ad-449b-b4f5-6bf4735fc597",
          "game_id": "m0r0",
          "game_version": "492f87ba",
          "published_at": "2026-06-10T17:31:15.944941Z",
          "score": 4.761904761904762,
          "model": "gpt-5-6-sol",
          "runner": "benchmark_agent",
          "config": "openai-gpt-5-6-sol-high",
          "verified_publish": true,
          "levels_completed": 1,
          "actions": 584,
          "resets": 3,
          "state": "GAME_OVER",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            107.01545778834722,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            29,
            555,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            30,
            111,
            203,
            26,
            500,
            237
          ]
        },
        {
          "session_id": "3fb587aa-1eb9-4962-aae8-70ce03f2b2a7",
          "game_id": "m0r0",
          "game_version": "492f87ba",
          "published_at": "2026-06-10T10:26:49.382209Z",
          "score": 1.3190871916633686,
          "model": "gpt-5-6-sol",
          "runner": "benchmark_agent",
          "config": "openai-gpt-5-6-sol-medium",
          "verified_publish": true,
          "levels_completed": 1,
          "actions": 612,
          "resets": 3,
          "state": "GAME_OVER",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            27.700831024930743,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            57,
            555,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            30,
            111,
            203,
            26,
            500,
            237
          ]
        },
        {
          "session_id": "9698d275-dd3b-4b9c-b2c5-4e937366a6fa",
          "game_id": "m0r0",
          "game_version": "492f87ba",
          "published_at": "2026-06-09T23:30:23.487151Z",
          "score": 3.935458480913026,
          "model": "gpt-5-6-sol",
          "runner": "benchmark_agent",
          "config": "openai-gpt-5-6-sol-low",
          "verified_publish": true,
          "levels_completed": 1,
          "actions": 588,
          "resets": 3,
          "state": "GAME_OVER",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            82.64462809917354,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            33,
            555,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            30,
            111,
            203,
            26,
            500,
            237
          ]
        },
        {
          "session_id": "fad0ff63-6bd2-4562-a8f0-69866e309000",
          "game_id": "m0r0",
          "game_version": "492f87ba",
          "published_at": "2026-05-29T22:32:48.702252Z",
          "score": 3.7073652990608004,
          "model": "claude-opus-4-8",
          "runner": "benchmark_agent",
          "config": "anthropic-opus-4-8-high",
          "verified_publish": true,
          "levels_completed": 1,
          "actions": 287,
          "resets": 1,
          "state": "GAME_OVER",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            77.8546712802768,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            34,
            253,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            30,
            111,
            203,
            26,
            500,
            237
          ]
        },
        {
          "session_id": "a5d85819-1be3-4a6b-a2b9-6bb5789b68d3",
          "game_id": "m0r0",
          "game_version": "492f87ba",
          "published_at": "2026-05-01T16:15:04.201495Z",
          "score": 0,
          "model": "gpt-5.5-2026-04-23",
          "runner": "benchmark_agent",
          "config": "openai-gpt-5-5-2026-04-23-high",
          "verified_publish": true,
          "levels_completed": 0,
          "actions": 150,
          "resets": 1,
          "state": "NOT_FINISHED",
          "completed": false,
          "number_of_levels": 6,
          "level_scores": [
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            150,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            30,
            111,
            203,
            26,
            500,
            237
          ]
        },
        {
          "session_id": "ed297f4e-ca26-4300-8731-e0101006d62a",
          "game_id": "m0r0",
          "game_version": "492f87ba",
          "published_at": "2026-05-01T05:08:33.599362Z",
          "score": 0,
          "model": "claude-opus-4-7",
          "runner": "benchmark_agent",
          "config": "anthropic-opus-4-7-high",
          "verified_publish": true,
          "levels_completed": 0,
          "actions": 150,
          "resets": 3,
          "state": "NOT_FINISHED",
          "completed": false,
          "number_of_levels": 6,
          "level_scores": [
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            150,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            30,
            111,
            203,
            26,
            500,
            237
          ]
        },
        {
          "session_id": "35f68321-e665-48bb-9b29-c5bc1282234b",
          "game_id": "m0r0",
          "game_version": "492f87ba",
          "published_at": "2026-04-14T17:43:53.555239Z",
          "score": 0,
          "model": "grok-4.20-beta-0309-reasoning",
          "runner": "benchmark_agent",
          "config": "xai-grok-4-20-beta-0309-reasoning",
          "verified_publish": true,
          "levels_completed": 0,
          "actions": 56,
          "resets": 6,
          "state": "NOT_FINISHED",
          "completed": false,
          "number_of_levels": 6,
          "level_scores": [
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            56,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            30,
            111,
            203,
            26,
            500,
            237
          ]
        },
        {
          "session_id": "8fe28f67-034a-4797-94ed-6302f48e9827",
          "game_id": "m0r0",
          "game_version": "492f87ba",
          "published_at": "2026-04-14T17:43:53.260615Z",
          "score": 0,
          "model": "gpt-5.4-2026-03-05",
          "runner": "benchmark_agent",
          "config": "openai-gpt-5-4-2026-03-05-high",
          "verified_publish": true,
          "levels_completed": 0,
          "actions": 123,
          "resets": 0,
          "state": "NOT_FINISHED",
          "completed": false,
          "number_of_levels": 6,
          "level_scores": [
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            123,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            30,
            111,
            203,
            26,
            500,
            237
          ]
        },
        {
          "session_id": "aaa76ec5-ee97-468b-9079-10f918391592",
          "game_id": "m0r0",
          "game_version": "492f87ba",
          "published_at": "2026-04-14T17:43:52.833049Z",
          "score": 0,
          "model": "gemini-3.1-pro-preview",
          "runner": "benchmark_agent",
          "config": "google-gemini-3-1-pro-preview",
          "verified_publish": true,
          "levels_completed": 0,
          "actions": 150,
          "resets": 0,
          "state": "NOT_FINISHED",
          "completed": false,
          "number_of_levels": 6,
          "level_scores": [
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            150,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            30,
            111,
            203,
            26,
            500,
            237
          ]
        },
        {
          "session_id": "2a76a069-ab80-4f1c-852f-440ffd1f9ba1",
          "game_id": "m0r0",
          "game_version": "492f87ba",
          "published_at": "2026-04-14T17:43:52.277704Z",
          "score": 0,
          "model": "claude-opus-4-6",
          "runner": "benchmark_agent",
          "config": "anthropic-opus-4-6-max-effort",
          "verified_publish": true,
          "levels_completed": 0,
          "actions": 150,
          "resets": 26,
          "state": "NOT_FINISHED",
          "completed": false,
          "number_of_levels": 6,
          "level_scores": [
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            150,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            30,
            111,
            203,
            26,
            500,
            237
          ]
        }
      ]
    },
    {
      "game": "tu93",
      "source": "https://arcprize.org/tasks/tu93",
      "runs": [
        {
          "session_id": "c644f00b-2d08-4675-8883-d8fb5233e3d0",
          "game_id": "tu93",
          "game_version": "0768757b",
          "published_at": "2026-09-25T06:55:07.830991Z",
          "score": 1.9388781718055104,
          "model": "gpt-6-sol",
          "runner": "benchmark_agent",
          "config": "openai-gpt-6-sol-none-provider-adapter",
          "verified_publish": true,
          "levels_completed": 2,
          "actions": 236,
          "resets": 39,
          "state": "GAME_OVER",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            26.36961285609934,
            30.439952437574313,
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            37,
            29,
            170,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            19,
            16,
            34,
            42,
            123,
            80,
            14,
            23,
            111
          ]
        },
        {
          "session_id": "21014b75-66c2-40cc-857f-e963df26a647",
          "game_id": "tu93",
          "game_version": "0768757b",
          "published_at": "2026-09-25T06:26:58.155552Z",
          "score": 2.98558724076409,
          "model": "gpt-6-sol",
          "runner": "benchmark_agent",
          "config": "openai-gpt-6-sol-low-provider-adapter",
          "verified_publish": true,
          "levels_completed": 2,
          "actions": 224,
          "resets": 25,
          "state": "GAME_OVER",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            37.56503642039542,
            48.39319470699432,
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            31,
            23,
            170,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            19,
            16,
            34,
            42,
            123,
            80,
            14,
            23,
            111
          ]
        },
        {
          "session_id": "5cca49f6-d102-4b97-8946-c8e3cb6497c1",
          "game_id": "tu93",
          "game_version": "0768757b",
          "published_at": "2026-09-25T06:12:16.66Z",
          "score": 0,
          "model": "gpt-6-sol",
          "runner": "benchmark_agent",
          "config": "openai-gpt-6-sol-none",
          "verified_publish": true,
          "levels_completed": 0,
          "actions": 95,
          "resets": 1,
          "state": "GAME_OVER",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            0,
            0,
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            95,
            0,
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            19,
            16,
            34,
            42,
            123,
            80,
            14,
            23,
            111
          ]
        },
        {
          "session_id": "19feccb9-7078-4a94-9c77-e81965806e33",
          "game_id": "tu93",
          "game_version": "0768757b",
          "published_at": "2026-09-25T05:56:28.340483Z",
          "score": 5.58422090729783,
          "model": "gpt-6-sol",
          "runner": "benchmark_agent",
          "config": "openai-gpt-6-sol-medium-provider-adapter",
          "verified_publish": true,
          "levels_completed": 2,
          "actions": 224,
          "resets": 23,
          "state": "GAME_OVER",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            23.73438527284681,
            113.77777777777777,
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            39,
            15,
            170,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            19,
            16,
            34,
            42,
            123,
            80,
            14,
            23,
            111
          ]
        },
        {
          "session_id": "eac29481-890f-4cfb-afc9-7e373863053a",
          "game_id": "tu93",
          "game_version": "0768757b",
          "published_at": "2026-09-25T05:50:15.443012Z",
          "score": 0,
          "model": "gpt-6-sol",
          "runner": "benchmark_agent",
          "config": "openai-gpt-6-sol-low",
          "verified_publish": true,
          "levels_completed": 0,
          "actions": 95,
          "resets": 1,
          "state": "GAME_OVER",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            0,
            0,
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            95,
            0,
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            19,
            16,
            34,
            42,
            123,
            80,
            14,
            23,
            111
          ]
        },
        {
          "session_id": "3f143116-a535-454c-a734-7b1aa99bf631",
          "game_id": "tu93",
          "game_version": "0768757b",
          "published_at": "2026-09-25T05:12:06.092141Z",
          "score": 0,
          "model": "gpt-6-sol",
          "runner": "benchmark_agent",
          "config": "openai-gpt-6-sol-medium",
          "verified_publish": true,
          "levels_completed": 0,
          "actions": 95,
          "resets": 1,
          "state": "GAME_OVER",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            0,
            0,
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            95,
            0,
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            19,
            16,
            34,
            42,
            123,
            80,
            14,
            23,
            111
          ]
        },
        {
          "session_id": "7ed4003b-6084-4ece-85ac-68cb73a8a371",
          "game_id": "tu93",
          "game_version": "0768757b",
          "published_at": "2026-09-25T04:57:38.799668Z",
          "score": 3.7944066515495085,
          "model": "gpt-6-sol",
          "runner": "benchmark_agent",
          "config": "openai-gpt-6-sol-high-provider-adapter",
          "verified_publish": true,
          "levels_completed": 2,
          "actions": 215,
          "resets": 23,
          "state": "GAME_OVER",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            81.859410430839,
            44.44444444444444,
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            21,
            24,
            170,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            19,
            16,
            34,
            42,
            123,
            80,
            14,
            23,
            111
          ]
        },
        {
          "session_id": "7f5bb895-401e-4fd1-bb1f-92d818f42e8f",
          "game_id": "tu93",
          "game_version": "0768757b",
          "published_at": "2026-09-25T04:46:58.838827Z",
          "score": 0,
          "model": "gpt-6-sol",
          "runner": "benchmark_agent",
          "config": "openai-gpt-6-sol-max",
          "verified_publish": true,
          "levels_completed": 0,
          "actions": 95,
          "resets": 2,
          "state": "GAME_OVER",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            0,
            0,
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            95,
            0,
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            19,
            16,
            34,
            42,
            123,
            80,
            14,
            23,
            111
          ]
        },
        {
          "session_id": "ae8e16b0-005f-45a8-b8a4-c7ac3bae0e09",
          "game_id": "tu93",
          "game_version": "0768757b",
          "published_at": "2026-09-25T04:39:54.646225Z",
          "score": 2.4766975308641976,
          "model": "gpt-6-sol",
          "runner": "benchmark_agent",
          "config": "openai-gpt-6-sol-xhigh-provider-adapter",
          "verified_publish": true,
          "levels_completed": 2,
          "actions": 234,
          "resets": 28,
          "state": "GAME_OVER",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            22.5625,
            44.44444444444444,
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            40,
            24,
            170,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            19,
            16,
            34,
            42,
            123,
            80,
            14,
            23,
            111
          ]
        },
        {
          "session_id": "dc7f93ca-676e-4e29-ba7f-bb9222d59812",
          "game_id": "tu93",
          "game_version": "0768757b",
          "published_at": "2026-09-25T04:31:39.294739Z",
          "score": 0,
          "model": "gpt-6-sol",
          "runner": "benchmark_agent",
          "config": "openai-gpt-6-sol-high",
          "verified_publish": true,
          "levels_completed": 0,
          "actions": 95,
          "resets": 1,
          "state": "NOT_FINISHED",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            0,
            0,
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            95,
            0,
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            19,
            16,
            34,
            42,
            123,
            80,
            14,
            23,
            111
          ]
        },
        {
          "session_id": "747f57af-493d-4066-b0a2-d34eb9b6bf23",
          "game_id": "tu93",
          "game_version": "0768757b",
          "published_at": "2026-09-25T04:16:08.764773Z",
          "score": 0.43386815696172104,
          "model": "gpt-6-sol",
          "runner": "benchmark_agent",
          "config": "openai-gpt-6-sol-xhigh",
          "verified_publish": true,
          "levels_completed": 1,
          "actions": 123,
          "resets": 5,
          "state": "GAME_OVER",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            19.524067063277446,
            0,
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            43,
            80,
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            19,
            16,
            34,
            42,
            123,
            80,
            14,
            23,
            111
          ]
        },
        {
          "session_id": "b4375afd-bac6-4d8d-8c0e-00840a824feb",
          "game_id": "tu93",
          "game_version": "0768757b",
          "published_at": "2026-09-21T17:29:33.641944Z",
          "score": 16.009694827964058,
          "model": "gemini-3.8-flash",
          "runner": "benchmark_agent",
          "config": "google-gemini-3-8-flash-low-provider-adapter",
          "verified_publish": true,
          "levels_completed": 4,
          "actions": 764,
          "resets": 81,
          "state": "GAME_OVER",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            62.67361111111111,
            113.77777777777777,
            115,
            21.301775147928996,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            24,
            15,
            19,
            91,
            615,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            19,
            16,
            34,
            42,
            123,
            80,
            14,
            23,
            111
          ]
        },
        {
          "session_id": "ab69271d-0626-4641-9874-6bc1629330d6",
          "game_id": "tu93",
          "game_version": "0768757b",
          "published_at": "2026-09-20T14:33:57.13834Z",
          "score": 22.22222222222222,
          "model": "gemini-3.8-flash",
          "runner": "benchmark_agent",
          "config": "google-gemini-3-8-flash-medium-provider-adapter",
          "verified_publish": true,
          "levels_completed": 4,
          "actions": 709,
          "resets": 33,
          "state": "GAME_OVER",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            111.41975308641976,
            113.77777777777777,
            115,
            115,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            18,
            15,
            24,
            37,
            615,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            19,
            16,
            34,
            42,
            123,
            80,
            14,
            23,
            111
          ]
        },
        {
          "session_id": "4995f6a1-1cff-4695-b351-2d0152b6d2ec",
          "game_id": "tu93",
          "game_version": "0768757b",
          "published_at": "2026-09-18T18:15:47.211618Z",
          "score": 33.33333333333333,
          "model": "gemini-3.8-flash",
          "runner": "benchmark_agent",
          "config": "google-gemini-3-8-flash-high-provider-adapter",
          "verified_publish": true,
          "levels_completed": 5,
          "actions": 522,
          "resets": 38,
          "state": "GAME_OVER",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            111.41975308641976,
            115,
            115,
            110.25,
            115,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            18,
            10,
            19,
            40,
            35,
            400,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            19,
            16,
            34,
            42,
            123,
            80,
            14,
            23,
            111
          ]
        },
        {
          "session_id": "15cd5794-a190-4b76-a115-de14067e3b5a",
          "game_id": "tu93",
          "game_version": "0768757b",
          "published_at": "2026-09-16T19:26:40.994846Z",
          "score": 0,
          "model": "gpt-6-luna",
          "runner": "benchmark_agent",
          "config": "openai-gpt-6-luna-none-provider-adapter",
          "verified_publish": true,
          "levels_completed": 0,
          "actions": 95,
          "resets": 1,
          "state": "GAME_OVER",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            0,
            0,
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            95,
            0,
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            19,
            16,
            34,
            42,
            123,
            80,
            14,
            23,
            111
          ]
        },
        {
          "session_id": "58dd3634-ac20-4bd3-aa47-94c60494b7a4",
          "game_id": "tu93",
          "game_version": "0768757b",
          "published_at": "2026-09-16T19:22:09.547864Z",
          "score": 0,
          "model": "gpt-6-luna",
          "runner": "benchmark_agent",
          "config": "openai-gpt-6-luna-none",
          "verified_publish": true,
          "levels_completed": 0,
          "actions": 95,
          "resets": 1,
          "state": "GAME_OVER",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            0,
            0,
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            95,
            0,
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            19,
            16,
            34,
            42,
            123,
            80,
            14,
            23,
            111
          ]
        },
        {
          "session_id": "62cc0288-b529-46e1-9fdb-5adf1e2e97c4",
          "game_id": "tu93",
          "game_version": "0768757b",
          "published_at": "2026-09-16T15:59:12.997643Z",
          "score": 0,
          "model": "gpt-6-luna",
          "runner": "benchmark_agent",
          "config": "openai-gpt-6-luna-xhigh",
          "verified_publish": true,
          "levels_completed": 0,
          "actions": 95,
          "resets": 2,
          "state": "GAME_OVER",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            0,
            0,
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            95,
            0,
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            19,
            16,
            34,
            42,
            123,
            80,
            14,
            23,
            111
          ]
        },
        {
          "session_id": "c7ca1647-dc1e-47dc-b9fe-03badd8d61f5",
          "game_id": "tu93",
          "game_version": "0768757b",
          "published_at": "2026-09-16T15:49:36.60544Z",
          "score": 0,
          "model": "gpt-6-luna",
          "runner": "benchmark_agent",
          "config": "openai-gpt-6-luna-high-provider-adapter",
          "verified_publish": true,
          "levels_completed": 0,
          "actions": 95,
          "resets": 3,
          "state": "GAME_OVER",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            0,
            0,
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            95,
            0,
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            19,
            16,
            34,
            42,
            123,
            80,
            14,
            23,
            111
          ]
        },
        {
          "session_id": "94828805-da5b-422b-b2f8-816582308fd5",
          "game_id": "tu93",
          "game_version": "0768757b",
          "published_at": "2026-09-16T14:45:45.340924Z",
          "score": 2.2222222222222223,
          "model": "gpt-6-1-sol",
          "runner": "benchmark_agent",
          "config": "openai-gpt-6-1-sol-low",
          "verified_publish": true,
          "levels_completed": 1,
          "actions": 77,
          "resets": 11,
          "state": "GAME_OVER",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            111.41975308641976,
            0,
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            18,
            59,
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            19,
            16,
            34,
            42,
            123,
            80,
            14,
            23,
            111
          ]
        },
        {
          "session_id": "ebb66304-7ad0-48f3-a791-bdb0567bb03b",
          "game_id": "tu93",
          "game_version": "0768757b",
          "published_at": "2026-09-16T09:16:50.510756Z",
          "score": 0,
          "model": "gpt-6-luna",
          "runner": "benchmark_agent",
          "config": "openai-gpt-6-luna-low-provider-adapter",
          "verified_publish": true,
          "levels_completed": 0,
          "actions": 95,
          "resets": 1,
          "state": "GAME_OVER",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            0,
            0,
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            95,
            0,
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            19,
            16,
            34,
            42,
            123,
            80,
            14,
            23,
            111
          ]
        },
        {
          "session_id": "b91e819a-c1a0-42cd-afc5-26119ef55f52",
          "game_id": "tu93",
          "game_version": "0768757b",
          "published_at": "2026-09-16T09:08:39.530629Z",
          "score": 0,
          "model": "gpt-6-luna",
          "runner": "benchmark_agent",
          "config": "openai-gpt-6-luna-low",
          "verified_publish": true,
          "levels_completed": 0,
          "actions": 93,
          "resets": 2,
          "state": "GAME_OVER",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            0,
            0,
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            93,
            0,
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            19,
            16,
            34,
            42,
            123,
            80,
            14,
            23,
            111
          ]
        },
        {
          "session_id": "d3554e81-a04c-4168-a9f8-5fe44488fba5",
          "game_id": "tu93",
          "game_version": "0768757b",
          "published_at": "2026-09-16T09:04:38.527099Z",
          "score": 0,
          "model": "gpt-6-luna",
          "runner": "benchmark_agent",
          "config": "openai-gpt-6-luna-medium-provider-adapter",
          "verified_publish": true,
          "levels_completed": 0,
          "actions": 95,
          "resets": 6,
          "state": "NOT_FINISHED",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            0,
            0,
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            95,
            0,
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            19,
            16,
            34,
            42,
            123,
            80,
            14,
            23,
            111
          ]
        },
        {
          "session_id": "4956aaed-15d8-4c3b-96ba-d78fdc89215e",
          "game_id": "tu93",
          "game_version": "0768757b",
          "published_at": "2026-09-16T08:47:53.133525Z",
          "score": 0,
          "model": "gpt-6-luna",
          "runner": "benchmark_agent",
          "config": "openai-gpt-6-luna-medium",
          "verified_publish": true,
          "levels_completed": 0,
          "actions": 95,
          "resets": 1,
          "state": "NOT_FINISHED",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            0,
            0,
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            95,
            0,
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            19,
            16,
            34,
            42,
            123,
            80,
            14,
            23,
            111
          ]
        },
        {
          "session_id": "8a9abae0-d790-44b8-b370-0f362eac0c3a",
          "game_id": "tu93",
          "game_version": "0768757b",
          "published_at": "2026-09-16T08:25:14.667727Z",
          "score": 0,
          "model": "gpt-6-luna",
          "runner": "benchmark_agent",
          "config": "openai-gpt-6-luna-high",
          "verified_publish": true,
          "levels_completed": 0,
          "actions": 95,
          "resets": 3,
          "state": "GAME_OVER",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            0,
            0,
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            95,
            0,
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            19,
            16,
            34,
            42,
            123,
            80,
            14,
            23,
            111
          ]
        },
        {
          "session_id": "8215d819-4b54-4b15-bbca-735475e153b1",
          "game_id": "tu93",
          "game_version": "0768757b",
          "published_at": "2026-09-16T07:42:00.458125Z",
          "score": 0,
          "model": "gpt-6-luna",
          "runner": "benchmark_agent",
          "config": "openai-gpt-6-luna-xhigh-provider-adapter",
          "verified_publish": true,
          "levels_completed": 0,
          "actions": 95,
          "resets": 3,
          "state": "NOT_FINISHED",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            0,
            0,
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            95,
            0,
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            19,
            16,
            34,
            42,
            123,
            80,
            14,
            23,
            111
          ]
        },
        {
          "session_id": "043b0e33-0543-4277-8956-4f2a960d112b",
          "game_id": "tu93",
          "game_version": "0768757b",
          "published_at": "2026-09-16T06:48:52.999785Z",
          "score": 0,
          "model": "gpt-6-luna",
          "runner": "benchmark_agent",
          "config": "openai-gpt-6-luna-max",
          "verified_publish": true,
          "levels_completed": 0,
          "actions": 95,
          "resets": 3,
          "state": "NOT_FINISHED",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            0,
            0,
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            95,
            0,
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            19,
            16,
            34,
            42,
            123,
            80,
            14,
            23,
            111
          ]
        },
        {
          "session_id": "85f53be4-0c2e-4b0b-8487-11fb9e10b04e",
          "game_id": "tu93",
          "game_version": "0768757b",
          "published_at": "2026-09-16T06:42:43.189299Z",
          "score": 0,
          "model": "gpt-6-luna",
          "runner": "benchmark_agent",
          "config": "openai-gpt-6-luna-max-provider-adapter",
          "verified_publish": true,
          "levels_completed": 0,
          "actions": 95,
          "resets": 1,
          "state": "GAME_OVER",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            0,
            0,
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            95,
            0,
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            19,
            16,
            34,
            42,
            123,
            80,
            14,
            23,
            111
          ]
        },
        {
          "session_id": "7f8dce9e-4ec9-4851-8e9a-f65f1397cab7",
          "game_id": "tu93",
          "game_version": "0768757b",
          "published_at": "2026-09-16T05:20:07.243019Z",
          "score": 100,
          "model": "gpt-6-1-sol",
          "runner": "benchmark_agent",
          "config": "openai-gpt-6-1-sol-high-provider-adapter",
          "verified_publish": true,
          "levels_completed": 9,
          "actions": 218,
          "resets": 3,
          "state": "WIN",
          "completed": true,
          "number_of_levels": 0,
          "level_scores": [
            111.41975308641976,
            115,
            115,
            115,
            115,
            115,
            100,
            100,
            115
          ],
          "level_actions": [
            18,
            10,
            27,
            20,
            49,
            28,
            14,
            23,
            29
          ],
          "level_baseline_actions": [
            19,
            16,
            34,
            42,
            123,
            80,
            14,
            23,
            111
          ]
        },
        {
          "session_id": "218c769a-d570-4488-9864-664e451c7fa1",
          "game_id": "tu93",
          "game_version": "0768757b",
          "published_at": "2026-09-16T05:05:32.511375Z",
          "score": 100,
          "model": "gpt-6-1-sol",
          "runner": "benchmark_agent",
          "config": "openai-gpt-6-1-sol-max-provider-adapter",
          "verified_publish": true,
          "levels_completed": 9,
          "actions": 217,
          "resets": 2,
          "state": "WIN",
          "completed": true,
          "number_of_levels": 0,
          "level_scores": [
            111.41975308641976,
            115,
            115,
            115,
            115,
            115,
            100,
            115,
            115
          ],
          "level_actions": [
            18,
            10,
            27,
            19,
            29,
            30,
            14,
            21,
            49
          ],
          "level_baseline_actions": [
            19,
            16,
            34,
            42,
            123,
            80,
            14,
            23,
            111
          ]
        },
        {
          "session_id": "e6cde68b-aa13-4bb0-8713-c30df4273dcf",
          "game_id": "tu93",
          "game_version": "0768757b",
          "published_at": "2026-09-16T05:04:14.149722Z",
          "score": 0.11930729063388196,
          "model": "gpt-6-1-sol",
          "runner": "benchmark_agent",
          "config": "openai-gpt-6-1-sol-medium",
          "verified_publish": true,
          "levels_completed": 1,
          "actions": 162,
          "resets": 15,
          "state": "GAME_OVER",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            5.368828078524688,
            0,
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            82,
            80,
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            19,
            16,
            34,
            42,
            123,
            80,
            14,
            23,
            111
          ]
        },
        {
          "session_id": "b4896e52-a8a6-4d4b-97ad-fba921c5ec81",
          "game_id": "tu93",
          "game_version": "0768757b",
          "published_at": "2026-09-16T04:15:06.291803Z",
          "score": 100,
          "model": "gpt-6-1-sol",
          "runner": "benchmark_agent",
          "config": "openai-gpt-6-1-sol-xhigh-provider-adapter",
          "verified_publish": true,
          "levels_completed": 9,
          "actions": 275,
          "resets": 6,
          "state": "WIN",
          "completed": true,
          "number_of_levels": 0,
          "level_scores": [
            111.41975308641976,
            115,
            115,
            115,
            115,
            115,
            34.027777777777786,
            115,
            115
          ],
          "level_actions": [
            18,
            10,
            27,
            18,
            29,
            38,
            24,
            21,
            90
          ],
          "level_baseline_actions": [
            19,
            16,
            34,
            42,
            123,
            80,
            14,
            23,
            111
          ]
        },
        {
          "session_id": "7b512b97-3f42-41d0-94fe-02bd773aed37",
          "game_id": "tu93",
          "game_version": "0768757b",
          "published_at": "2026-09-16T03:39:32.030235Z",
          "score": 2.0055555555555555,
          "model": "gpt-6-1-sol",
          "runner": "benchmark_agent",
          "config": "openai-gpt-6-1-sol-high",
          "verified_publish": true,
          "levels_completed": 1,
          "actions": 100,
          "resets": 14,
          "state": "GAME_OVER",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            90.25,
            0,
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            20,
            80,
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            19,
            16,
            34,
            42,
            123,
            80,
            14,
            23,
            111
          ]
        },
        {
          "session_id": "3f7af87e-66da-4082-a099-3f9f4c3a5e73",
          "game_id": "tu93",
          "game_version": "0768757b",
          "published_at": "2026-09-16T03:32:28.705151Z",
          "score": 2.2222222222222223,
          "model": "gpt-6-1-sol",
          "runner": "benchmark_agent",
          "config": "openai-gpt-6-1-sol-max",
          "verified_publish": true,
          "levels_completed": 1,
          "actions": 67,
          "resets": 8,
          "state": "GAME_OVER",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            111.41975308641976,
            0,
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            18,
            49,
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            19,
            16,
            34,
            42,
            123,
            80,
            14,
            23,
            111
          ]
        },
        {
          "session_id": "1114dfd8-6e39-434f-b5b9-0ae822fc1cbb",
          "game_id": "tu93",
          "game_version": "0768757b",
          "published_at": "2026-09-16T03:24:50.559649Z",
          "score": 2.2222222222222223,
          "model": "gpt-6-1-sol",
          "runner": "benchmark_agent",
          "config": "openai-gpt-6-1-sol-xhigh",
          "verified_publish": true,
          "levels_completed": 1,
          "actions": 98,
          "resets": 15,
          "state": "GAME_OVER",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            111.41975308641976,
            0,
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            18,
            80,
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            19,
            16,
            34,
            42,
            123,
            80,
            14,
            23,
            111
          ]
        },
        {
          "session_id": "27adb99f-c5be-43b6-ad9a-e9ffe6a838e2",
          "game_id": "tu93",
          "game_version": "0768757b",
          "published_at": "2026-09-09T19:45:39.027742Z",
          "score": 4.373030875866414,
          "model": "gemini-3.8-flash",
          "runner": "benchmark_agent",
          "config": "google-gemini-3-8-flash-high",
          "verified_publish": true,
          "levels_completed": 2,
          "actions": 212,
          "resets": 10,
          "state": "GAME_OVER",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            100,
            48.39319470699432,
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            19,
            23,
            170,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            19,
            16,
            34,
            42,
            123,
            80,
            14,
            23,
            111
          ]
        },
        {
          "session_id": "2bdf62b4-0940-4d26-bfdb-98d1b385cb47",
          "game_id": "tu93",
          "game_version": "0768757b",
          "published_at": "2026-09-09T18:17:37.598852Z",
          "score": 1.9751086749777358,
          "model": "gemini-3.8-flash",
          "runner": "benchmark_agent",
          "config": "google-gemini-3-8-flash-medium",
          "verified_publish": true,
          "levels_completed": 2,
          "actions": 241,
          "resets": 33,
          "state": "GAME_OVER",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            18.64669421487603,
            35.11659807956104,
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            44,
            27,
            170,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            19,
            16,
            34,
            42,
            123,
            80,
            14,
            23,
            111
          ]
        },
        {
          "session_id": "51f79ce3-abbf-4a84-8be8-312a6e6fd110",
          "game_id": "tu93",
          "game_version": "0768757b",
          "published_at": "2026-09-09T17:24:34.040068Z",
          "score": 0,
          "model": "gemini-3.8-flash",
          "runner": "benchmark_agent",
          "config": "google-gemini-3-8-flash-low",
          "verified_publish": true,
          "levels_completed": 0,
          "actions": 95,
          "resets": 1,
          "state": "GAME_OVER",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            0,
            0,
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            95,
            0,
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            19,
            16,
            34,
            42,
            123,
            80,
            14,
            23,
            111
          ]
        },
        {
          "session_id": "0e20aa49-a8ff-4c85-b469-1563db999ae8",
          "game_id": "tu93",
          "game_version": "0768757b",
          "published_at": "2026-08-11T06:23:25.417427Z",
          "score": 0,
          "model": "grok-4.6",
          "runner": "benchmark_agent",
          "config": "xai-grok-4-6-xhigh",
          "verified_publish": true,
          "levels_completed": 0,
          "actions": 95,
          "resets": 2,
          "state": "GAME_OVER",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            0,
            0,
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            95,
            0,
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            19,
            16,
            34,
            42,
            123,
            80,
            14,
            23,
            111
          ]
        },
        {
          "session_id": "fc41585f-6b3c-4d84-9248-10642cf766cf",
          "game_id": "tu93",
          "game_version": "0768757b",
          "published_at": "2026-07-21T03:20:48.281657Z",
          "score": 3.5662932479804907,
          "model": "claude-opus-5",
          "runner": "benchmark_agent",
          "config": "anthropic-claude-opus-5-high",
          "verified_publish": true,
          "levels_completed": 2,
          "actions": 217,
          "resets": 31,
          "state": "GAME_OVER",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            90.25,
            35.11659807956104,
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            20,
            27,
            170,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            19,
            16,
            34,
            42,
            123,
            80,
            14,
            23,
            111
          ]
        },
        {
          "session_id": "f7cb5d8f-6d00-4153-8133-bc9ee70db8c4",
          "game_id": "tu93",
          "game_version": "0768757b",
          "published_at": "2026-06-24T10:39:48.504565Z",
          "score": 0,
          "model": "gpt-5-6-luna",
          "runner": "benchmark_agent",
          "config": "openai-gpt-5-6-luna-low",
          "verified_publish": true,
          "levels_completed": 0,
          "actions": 95,
          "resets": 6,
          "state": "NOT_FINISHED",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            0,
            0,
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            95,
            0,
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            19,
            16,
            34,
            42,
            123,
            80,
            14,
            23,
            111
          ]
        },
        {
          "session_id": "d1960d8b-67f0-4769-a506-c158c2ad0b27",
          "game_id": "tu93",
          "game_version": "0768757b",
          "published_at": "2026-06-24T10:15:36.914257Z",
          "score": 0,
          "model": "gpt-5-6-luna",
          "runner": "benchmark_agent",
          "config": "openai-gpt-5-6-luna-medium",
          "verified_publish": true,
          "levels_completed": 0,
          "actions": 95,
          "resets": 7,
          "state": "NOT_FINISHED",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            0,
            0,
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            95,
            0,
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            19,
            16,
            34,
            42,
            123,
            80,
            14,
            23,
            111
          ]
        },
        {
          "session_id": "16e7c871-2ba9-4655-9a36-503158198a74",
          "game_id": "tu93",
          "game_version": "0768757b",
          "published_at": "2026-06-24T03:05:50.833066Z",
          "score": 0,
          "model": "gpt-5-6-luna",
          "runner": "benchmark_agent",
          "config": "openai-gpt-5-6-luna-high",
          "verified_publish": true,
          "levels_completed": 0,
          "actions": 95,
          "resets": 12,
          "state": "NOT_FINISHED",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            0,
            0,
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            95,
            0,
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            19,
            16,
            34,
            42,
            123,
            80,
            14,
            23,
            111
          ]
        },
        {
          "session_id": "cf00dfe1-0d52-4ae4-9351-10ef4c32de39",
          "game_id": "tu93",
          "game_version": "0768757b",
          "published_at": "2026-06-23T23:17:13.863741Z",
          "score": 0,
          "model": "gpt-5-6-luna",
          "runner": "benchmark_agent",
          "config": "openai-gpt-5-6-luna-xhigh",
          "verified_publish": true,
          "levels_completed": 0,
          "actions": 95,
          "resets": 13,
          "state": "NOT_FINISHED",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            0,
            0,
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            95,
            0,
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            19,
            16,
            34,
            42,
            123,
            80,
            14,
            23,
            111
          ]
        },
        {
          "session_id": "373134ee-8bd7-4361-9ef9-a8138b9f323b",
          "game_id": "tu93",
          "game_version": "0768757b",
          "published_at": "2026-06-23T02:43:34.938329Z",
          "score": 0,
          "model": "gpt-5-6-luna",
          "runner": "benchmark_agent",
          "config": "openai-gpt-5-6-luna-max",
          "verified_publish": true,
          "levels_completed": 0,
          "actions": 95,
          "resets": 17,
          "state": "NOT_FINISHED",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            0,
            0,
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            95,
            0,
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            19,
            16,
            34,
            42,
            123,
            80,
            14,
            23,
            111
          ]
        },
        {
          "session_id": "01d1e1d4-c082-4fae-856a-37100c472e96",
          "game_id": "tu93",
          "game_version": "0768757b",
          "published_at": "2026-06-20T11:48:17.263493Z",
          "score": 0,
          "model": "gpt-5-6-terra",
          "runner": "benchmark_agent",
          "config": "openai-gpt-5-6-terra-low",
          "verified_publish": true,
          "levels_completed": 0,
          "actions": 95,
          "resets": 3,
          "state": "GAME_OVER",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            0,
            0,
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            95,
            0,
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            19,
            16,
            34,
            42,
            123,
            80,
            14,
            23,
            111
          ]
        },
        {
          "session_id": "733b3373-9006-4d7f-810a-0f6c6bf3d52d",
          "game_id": "tu93",
          "game_version": "0768757b",
          "published_at": "2026-06-20T08:59:17.928837Z",
          "score": 0,
          "model": "gpt-5-6-terra",
          "runner": "benchmark_agent",
          "config": "openai-gpt-5-6-terra-medium",
          "verified_publish": true,
          "levels_completed": 0,
          "actions": 95,
          "resets": 2,
          "state": "GAME_OVER",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            0,
            0,
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            95,
            0,
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            19,
            16,
            34,
            42,
            123,
            80,
            14,
            23,
            111
          ]
        },
        {
          "session_id": "12c75962-29b4-44e1-8578-c14a5c5a26ec",
          "game_id": "tu93",
          "game_version": "0768757b",
          "published_at": "2026-06-20T08:16:55.043677Z",
          "score": 0,
          "model": "gpt-5-6-terra",
          "runner": "benchmark_agent",
          "config": "openai-gpt-5-6-terra-high",
          "verified_publish": true,
          "levels_completed": 0,
          "actions": 95,
          "resets": 1,
          "state": "GAME_OVER",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            0,
            0,
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            95,
            0,
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            19,
            16,
            34,
            42,
            123,
            80,
            14,
            23,
            111
          ]
        },
        {
          "session_id": "686761de-5acb-483c-85db-6cf034a00017",
          "game_id": "tu93",
          "game_version": "0768757b",
          "published_at": "2026-06-20T00:38:23.410763Z",
          "score": 0,
          "model": "gpt-5-6-terra",
          "runner": "benchmark_agent",
          "config": "openai-gpt-5-6-terra-xhigh",
          "verified_publish": true,
          "levels_completed": 0,
          "actions": 95,
          "resets": 1,
          "state": "GAME_OVER",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            0,
            0,
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            95,
            0,
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            19,
            16,
            34,
            42,
            123,
            80,
            14,
            23,
            111
          ]
        },
        {
          "session_id": "b8113a26-308b-4860-afd2-530eaab3ef8a",
          "game_id": "tu93",
          "game_version": "0768757b",
          "published_at": "2026-06-18T21:44:04.652175Z",
          "score": 0,
          "model": "gpt-5-6-terra",
          "runner": "benchmark_agent",
          "config": "openai-gpt-5-6-terra-max",
          "verified_publish": true,
          "levels_completed": 0,
          "actions": 95,
          "resets": 1,
          "state": "GAME_OVER",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            0,
            0,
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            95,
            0,
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            19,
            16,
            34,
            42,
            123,
            80,
            14,
            23,
            111
          ]
        },
        {
          "session_id": "8456ed6d-3170-4075-99a9-9a9f613042e5",
          "game_id": "tu93",
          "game_version": "0768757b",
          "published_at": "2026-06-13T04:54:51.65112Z",
          "score": 1.3082173487022783,
          "model": "gpt-5-6-sol",
          "runner": "benchmark_agent",
          "config": "openai-gpt-5-6-sol-max",
          "verified_publish": true,
          "levels_completed": 2,
          "actions": 271,
          "resets": 34,
          "state": "GAME_OVER",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            49.51989026063101,
            4.6749452154857565,
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            27,
            74,
            170,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            19,
            16,
            34,
            42,
            123,
            80,
            14,
            23,
            111
          ]
        },
        {
          "session_id": "8eb2957d-33da-4db2-bfa7-e98947f594ff",
          "game_id": "tu93",
          "game_version": "0768757b",
          "published_at": "2026-06-11T15:34:33.524294Z",
          "score": 1.17959564704579,
          "model": "gpt-5-6-sol",
          "runner": "benchmark_agent",
          "config": "openai-gpt-5-6-sol-xhigh",
          "verified_publish": true,
          "levels_completed": 2,
          "actions": 270,
          "resets": 39,
          "state": "GAME_OVER",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            42.92508917954815,
            5.078357468756199,
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            29,
            71,
            170,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            19,
            16,
            34,
            42,
            123,
            80,
            14,
            23,
            111
          ]
        },
        {
          "session_id": "982b6849-0820-4574-a29a-163386709877",
          "game_id": "tu93",
          "game_version": "0768757b",
          "published_at": "2026-06-10T12:32:41.818968Z",
          "score": 1.593864449483497,
          "model": "gpt-5-6-sol",
          "runner": "benchmark_agent",
          "config": "openai-gpt-5-6-sol-high",
          "verified_publish": true,
          "levels_completed": 2,
          "actions": 273,
          "resets": 32,
          "state": "GAME_OVER",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            6.417777777777779,
            32.6530612244898,
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            75,
            28,
            170,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            19,
            16,
            34,
            42,
            123,
            80,
            14,
            23,
            111
          ]
        },
        {
          "session_id": "ae311391-f924-45f1-b56b-6a5f44ab4dca",
          "game_id": "tu93",
          "game_version": "0768757b",
          "published_at": "2026-06-10T07:45:48.195939Z",
          "score": 0,
          "model": "gpt-5-6-sol",
          "runner": "benchmark_agent",
          "config": "openai-gpt-5-6-sol-medium",
          "verified_publish": true,
          "levels_completed": 0,
          "actions": 95,
          "resets": 1,
          "state": "GAME_OVER",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            0,
            0,
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            95,
            0,
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            19,
            16,
            34,
            42,
            123,
            80,
            14,
            23,
            111
          ]
        },
        {
          "session_id": "03343ec5-6a6a-4253-9223-7a4119458e13",
          "game_id": "tu93",
          "game_version": "0768757b",
          "published_at": "2026-06-09T21:29:23.354769Z",
          "score": 0,
          "model": "gpt-5-6-sol",
          "runner": "benchmark_agent",
          "config": "openai-gpt-5-6-sol-low",
          "verified_publish": true,
          "levels_completed": 0,
          "actions": 95,
          "resets": 1,
          "state": "GAME_OVER",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            0,
            0,
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            95,
            0,
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            19,
            16,
            34,
            42,
            123,
            80,
            14,
            23,
            111
          ]
        },
        {
          "session_id": "59fce9ca-d398-43e3-bcd7-124df92642e4",
          "game_id": "tu93",
          "game_version": "0768757b",
          "published_at": "2026-05-29T17:51:09.798514Z",
          "score": 1.5164881327452215,
          "model": "claude-opus-4-8",
          "runner": "benchmark_agent",
          "config": "anthropic-opus-4-8-high",
          "verified_publish": true,
          "levels_completed": 1,
          "actions": 103,
          "resets": 13,
          "state": "GAME_OVER",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            68.24196597353497,
            0,
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            23,
            80,
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            19,
            16,
            34,
            42,
            123,
            80,
            14,
            23,
            111
          ]
        },
        {
          "session_id": "3beb8811-6f2a-4ce1-a1b6-bf53d1b4ad1d",
          "game_id": "tu93",
          "game_version": "0768757b",
          "published_at": "2026-05-01T05:08:42.421748Z",
          "score": 0,
          "model": "claude-opus-4-7",
          "runner": "benchmark_agent",
          "config": "anthropic-opus-4-7-high",
          "verified_publish": true,
          "levels_completed": 0,
          "actions": 95,
          "resets": 2,
          "state": "GAME_OVER",
          "completed": false,
          "number_of_levels": 9,
          "level_scores": [
            0,
            0,
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            95,
            0,
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            19,
            16,
            34,
            42,
            123,
            80,
            14,
            23,
            111
          ]
        },
        {
          "session_id": "f996a072-f821-446b-b8df-7bb5a5c57e46",
          "game_id": "tu93",
          "game_version": "0768757b",
          "published_at": "2026-04-30T23:23:36.75848Z",
          "score": 0,
          "model": "gpt-5.5-2026-04-23",
          "runner": "benchmark_agent",
          "config": "openai-gpt-5-5-2026-04-23-high",
          "verified_publish": true,
          "levels_completed": 0,
          "actions": 95,
          "resets": 1,
          "state": "GAME_OVER",
          "completed": false,
          "number_of_levels": 9,
          "level_scores": [
            0,
            0,
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            95,
            0,
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            19,
            16,
            34,
            42,
            123,
            80,
            14,
            23,
            111
          ]
        },
        {
          "session_id": "a401aab6-6787-4ddc-863d-ce6eecd26c12",
          "game_id": "tu93",
          "game_version": "0768757b",
          "published_at": "2026-04-14T17:44:06.651196Z",
          "score": 0,
          "model": "grok-4.20-beta-0309-reasoning",
          "runner": "benchmark_agent",
          "config": "xai-grok-4-20-beta-0309-reasoning",
          "verified_publish": true,
          "levels_completed": 0,
          "actions": 95,
          "resets": 5,
          "state": "NOT_FINISHED",
          "completed": false,
          "number_of_levels": 9,
          "level_scores": [
            0,
            0,
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            95,
            0,
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            19,
            16,
            34,
            42,
            123,
            80,
            14,
            23,
            111
          ]
        },
        {
          "session_id": "29453810-4ec4-4819-bb4a-e1e7f34f067d",
          "game_id": "tu93",
          "game_version": "0768757b",
          "published_at": "2026-04-14T17:44:06.34328Z",
          "score": 0,
          "model": "gpt-5.4-2026-03-05",
          "runner": "benchmark_agent",
          "config": "openai-gpt-5-4-2026-03-05-high",
          "verified_publish": true,
          "levels_completed": 0,
          "actions": 95,
          "resets": 1,
          "state": "GAME_OVER",
          "completed": false,
          "number_of_levels": 9,
          "level_scores": [
            0,
            0,
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            95,
            0,
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            19,
            16,
            34,
            42,
            123,
            80,
            14,
            23,
            111
          ]
        },
        {
          "session_id": "de486b95-c629-4959-a023-f5452ed72e29",
          "game_id": "tu93",
          "game_version": "0768757b",
          "published_at": "2026-04-14T17:44:05.661074Z",
          "score": 0,
          "model": "gemini-3.1-pro-preview",
          "runner": "benchmark_agent",
          "config": "google-gemini-3-1-pro-preview",
          "verified_publish": true,
          "levels_completed": 0,
          "actions": 67,
          "resets": 3,
          "state": "GAME_OVER",
          "completed": false,
          "number_of_levels": 9,
          "level_scores": [
            0,
            0,
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            67,
            0,
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            19,
            16,
            34,
            42,
            123,
            80,
            14,
            23,
            111
          ]
        },
        {
          "session_id": "7a695d89-c3e8-43dd-8f52-f4e559877b17",
          "game_id": "tu93",
          "game_version": "0768757b",
          "published_at": "2026-04-14T17:44:05.357868Z",
          "score": 0,
          "model": "claude-opus-4-6",
          "runner": "benchmark_agent",
          "config": "anthropic-opus-4-6-max-effort",
          "verified_publish": true,
          "levels_completed": 0,
          "actions": 95,
          "resets": 1,
          "state": "GAME_OVER",
          "completed": false,
          "number_of_levels": 9,
          "level_scores": [
            0,
            0,
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            95,
            0,
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            19,
            16,
            34,
            42,
            123,
            80,
            14,
            23,
            111
          ]
        }
      ]
    },
    {
      "game": "sk48",
      "source": "https://arcprize.org/tasks/sk48",
      "runs": [
        {
          "session_id": "4d4f5cd6-a7c7-424e-babf-1eac0bdc6444",
          "game_id": "sk48",
          "game_version": "d8078629",
          "published_at": "2026-09-25T15:26:25.518982Z",
          "score": 0.21355601469237834,
          "model": "gpt-6-sol",
          "runner": "benchmark_agent",
          "config": "openai-gpt-6-sol-max",
          "verified_publish": true,
          "levels_completed": 1,
          "actions": 517,
          "resets": 1,
          "state": "NOT_FINISHED",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            7.68801652892562,
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            220,
            297,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            61,
            177,
            101,
            103,
            230,
            181,
            125,
            92
          ]
        },
        {
          "session_id": "a374f124-2ae5-48ae-97e7-0bd1bb7ea806",
          "game_id": "sk48",
          "game_version": "d8078629",
          "published_at": "2026-09-25T12:13:43.011753Z",
          "score": 0,
          "model": "gpt-6-sol",
          "runner": "benchmark_agent",
          "config": "openai-gpt-6-sol-none",
          "verified_publish": true,
          "levels_completed": 0,
          "actions": 305,
          "resets": 3,
          "state": "GAME_OVER",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            0,
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            305,
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            61,
            177,
            101,
            103,
            230,
            181,
            125,
            92
          ]
        },
        {
          "session_id": "211d49a0-2576-4a91-8560-45aedd013d00",
          "game_id": "sk48",
          "game_version": "d8078629",
          "published_at": "2026-09-25T09:46:28.574861Z",
          "score": 3.425989671260322,
          "model": "gpt-6-sol",
          "runner": "benchmark_agent",
          "config": "openai-gpt-6-sol-xhigh-provider-adapter",
          "verified_publish": true,
          "levels_completed": 2,
          "actions": 1317,
          "resets": 5,
          "state": "GAME_OVER",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            115,
            4.1678140826857915,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            28,
            867,
            422,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            61,
            177,
            101,
            103,
            230,
            181,
            125,
            92
          ]
        },
        {
          "session_id": "e9ade73b-4c06-4004-b101-7b357cea46af",
          "game_id": "sk48",
          "game_version": "d8078629",
          "published_at": "2026-09-25T08:10:27.236275Z",
          "score": 0.5932111519232731,
          "model": "gpt-6-sol",
          "runner": "benchmark_agent",
          "config": "openai-gpt-6-sol-medium-provider-adapter",
          "verified_publish": true,
          "levels_completed": 1,
          "actions": 1017,
          "resets": 3,
          "state": "GAME_OVER",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            21.355601469237833,
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            132,
            885,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            61,
            177,
            101,
            103,
            230,
            181,
            125,
            92
          ]
        },
        {
          "session_id": "55c5125e-f49c-4118-95f0-98509eb0d35b",
          "game_id": "sk48",
          "game_version": "d8078629",
          "published_at": "2026-09-25T07:55:31.91262Z",
          "score": 0.40375434027777773,
          "model": "gpt-6-sol",
          "runner": "benchmark_agent",
          "config": "openai-gpt-6-sol-none-provider-adapter",
          "verified_publish": true,
          "levels_completed": 1,
          "actions": 1045,
          "resets": 7,
          "state": "NOT_FINISHED",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            14.535156249999998,
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            160,
            885,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            61,
            177,
            101,
            103,
            230,
            181,
            125,
            92
          ]
        },
        {
          "session_id": "ce6cc0a3-72c1-40c4-9a57-a163f44103e6",
          "game_id": "sk48",
          "game_version": "d8078629",
          "published_at": "2026-09-25T07:53:04.472156Z",
          "score": 2.7777777777777777,
          "model": "gpt-6-sol",
          "runner": "benchmark_agent",
          "config": "openai-gpt-6-sol-high-provider-adapter",
          "verified_publish": true,
          "levels_completed": 1,
          "actions": 922,
          "resets": 5,
          "state": "GAME_OVER",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            115,
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            37,
            885,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            61,
            177,
            101,
            103,
            230,
            181,
            125,
            92
          ]
        },
        {
          "session_id": "28e860e8-4d16-4170-becc-ed6eed3904b8",
          "game_id": "sk48",
          "game_version": "d8078629",
          "published_at": "2026-09-25T06:33:49.625274Z",
          "score": 0,
          "model": "gpt-6-sol",
          "runner": "benchmark_agent",
          "config": "openai-gpt-6-sol-low-provider-adapter",
          "verified_publish": true,
          "levels_completed": 0,
          "actions": 305,
          "resets": 4,
          "state": "NOT_FINISHED",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            0,
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            305,
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            61,
            177,
            101,
            103,
            230,
            181,
            125,
            92
          ]
        },
        {
          "session_id": "f1109ae2-e128-4b7a-828e-8b4eb62ae0b6",
          "game_id": "sk48",
          "game_version": "d8078629",
          "published_at": "2026-09-25T06:23:13.342171Z",
          "score": 0,
          "model": "gpt-6-sol",
          "runner": "benchmark_agent",
          "config": "openai-gpt-6-sol-xhigh",
          "verified_publish": true,
          "levels_completed": 0,
          "actions": 305,
          "resets": 1,
          "state": "GAME_OVER",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            0,
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            305,
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            61,
            177,
            101,
            103,
            230,
            181,
            125,
            92
          ]
        },
        {
          "session_id": "8a3e4787-c470-40a0-9c42-9b98bd9ad0da",
          "game_id": "sk48",
          "game_version": "d8078629",
          "published_at": "2026-09-25T06:19:47.328809Z",
          "score": 0,
          "model": "gpt-6-sol",
          "runner": "benchmark_agent",
          "config": "openai-gpt-6-sol-low",
          "verified_publish": true,
          "levels_completed": 0,
          "actions": 305,
          "resets": 1,
          "state": "GAME_OVER",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            0,
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            305,
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            61,
            177,
            101,
            103,
            230,
            181,
            125,
            92
          ]
        },
        {
          "session_id": "b2f21ae9-b3d0-435b-9e48-473bb0a264c3",
          "game_id": "sk48",
          "game_version": "d8078629",
          "published_at": "2026-09-25T05:41:52.482162Z",
          "score": 0,
          "model": "gpt-6-sol",
          "runner": "benchmark_agent",
          "config": "openai-gpt-6-sol-medium",
          "verified_publish": true,
          "levels_completed": 0,
          "actions": 305,
          "resets": 1,
          "state": "NOT_FINISHED",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            0,
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            305,
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            61,
            177,
            101,
            103,
            230,
            181,
            125,
            92
          ]
        },
        {
          "session_id": "2b27b6a9-2cc5-403c-b22d-6e8bc5445354",
          "game_id": "sk48",
          "game_version": "d8078629",
          "published_at": "2026-09-25T05:35:20.466114Z",
          "score": 0,
          "model": "gpt-6-sol",
          "runner": "benchmark_agent",
          "config": "openai-gpt-6-sol-high",
          "verified_publish": true,
          "levels_completed": 0,
          "actions": 305,
          "resets": 1,
          "state": "GAME_OVER",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            0,
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            305,
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            61,
            177,
            101,
            103,
            230,
            181,
            125,
            92
          ]
        },
        {
          "session_id": "1421a2d5-39da-4506-944d-c683018d6af8",
          "game_id": "sk48",
          "game_version": "d8078629",
          "published_at": "2026-09-21T15:51:43.499194Z",
          "score": 3.5904004832005896,
          "model": "gemini-3.8-flash",
          "runner": "benchmark_agent",
          "config": "google-gemini-3-8-flash-low-provider-adapter",
          "verified_publish": true,
          "levels_completed": 2,
          "actions": 1192,
          "resets": 9,
          "state": "GAME_OVER",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            115,
            7.127208697610614,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            24,
            663,
            505,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            61,
            177,
            101,
            103,
            230,
            181,
            125,
            92
          ]
        },
        {
          "session_id": "64d2174b-ee83-4d67-b4d9-439ba020f434",
          "game_id": "sk48",
          "game_version": "d8078629",
          "published_at": "2026-09-20T13:09:54.94711Z",
          "score": 5.993940856083347,
          "model": "gemini-3.8-flash",
          "runner": "benchmark_agent",
          "config": "google-gemini-3-8-flash-medium-provider-adapter",
          "verified_publish": true,
          "levels_completed": 3,
          "actions": 1364,
          "resets": 25,
          "state": "NOT_FINISHED",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            115,
            8.447183867601735,
            27.96250102793234,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            49,
            609,
            191,
            515,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            61,
            177,
            101,
            103,
            230,
            181,
            125,
            92
          ]
        },
        {
          "session_id": "3607a002-d4c3-471c-bfcd-1ee42a3470e2",
          "game_id": "sk48",
          "game_version": "d8078629",
          "published_at": "2026-09-18T08:33:01.518665Z",
          "score": 15.13427302784571,
          "model": "gemini-3.8-flash",
          "runner": "benchmark_agent",
          "config": "google-gemini-3-8-flash-high-provider-adapter",
          "verified_publish": true,
          "levels_completed": 4,
          "actions": 1493,
          "resets": 21,
          "state": "GAME_OVER",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            115,
            115,
            52.79747425081518,
            10.360351562500002,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            26,
            121,
            139,
            320,
            887,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            61,
            177,
            101,
            103,
            230,
            181,
            125,
            92
          ]
        },
        {
          "session_id": "62f77d4e-4948-48c3-8f38-10d41fe2337d",
          "game_id": "sk48",
          "game_version": "d8078629",
          "published_at": "2026-09-16T22:28:23.898553Z",
          "score": 6.758048793249752,
          "model": "gpt-6-1-sol",
          "runner": "benchmark_agent",
          "config": "openai-gpt-6-1-sol-high",
          "verified_publish": true,
          "levels_completed": 2,
          "actions": 742,
          "resets": 9,
          "state": "NOT_FINISHED",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            115,
            64.14487827849553,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            16,
            221,
            505,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            61,
            177,
            101,
            103,
            230,
            181,
            125,
            92
          ]
        },
        {
          "session_id": "b4924fa9-a02d-49ce-a68c-47050d2447d4",
          "game_id": "sk48",
          "game_version": "d8078629",
          "published_at": "2026-09-16T19:40:21.589187Z",
          "score": 0,
          "model": "gpt-6-luna",
          "runner": "benchmark_agent",
          "config": "openai-gpt-6-luna-low",
          "verified_publish": true,
          "levels_completed": 0,
          "actions": 305,
          "resets": 1,
          "state": "NOT_FINISHED",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            0,
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            305,
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            61,
            177,
            101,
            103,
            230,
            181,
            125,
            92
          ]
        },
        {
          "session_id": "5701ffb8-2d12-4351-953a-e3a8746c3889",
          "game_id": "sk48",
          "game_version": "d8078629",
          "published_at": "2026-09-16T19:35:59.904943Z",
          "score": 0,
          "model": "gpt-6-luna",
          "runner": "benchmark_agent",
          "config": "openai-gpt-6-luna-low-provider-adapter",
          "verified_publish": true,
          "levels_completed": 0,
          "actions": 305,
          "resets": 1,
          "state": "GAME_OVER",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            0,
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            305,
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            61,
            177,
            101,
            103,
            230,
            181,
            125,
            92
          ]
        },
        {
          "session_id": "ed528acd-acc3-4686-9af0-c44adc5e6923",
          "game_id": "sk48",
          "game_version": "d8078629",
          "published_at": "2026-09-16T19:33:27.678858Z",
          "score": 0,
          "model": "gpt-6-luna",
          "runner": "benchmark_agent",
          "config": "openai-gpt-6-luna-none-provider-adapter",
          "verified_publish": true,
          "levels_completed": 0,
          "actions": 305,
          "resets": 0,
          "state": "NOT_FINISHED",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            0,
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            305,
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            61,
            177,
            101,
            103,
            230,
            181,
            125,
            92
          ]
        },
        {
          "session_id": "5a168807-62fb-4416-b961-5bae55147bd5",
          "game_id": "sk48",
          "game_version": "d8078629",
          "published_at": "2026-09-16T19:29:32.484605Z",
          "score": 0,
          "model": "gpt-6-luna",
          "runner": "benchmark_agent",
          "config": "openai-gpt-6-luna-none",
          "verified_publish": true,
          "levels_completed": 0,
          "actions": 305,
          "resets": 8,
          "state": "NOT_FINISHED",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            0,
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            305,
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            61,
            177,
            101,
            103,
            230,
            181,
            125,
            92
          ]
        },
        {
          "session_id": "afe3ecd9-da7f-4452-b500-4eadc1c78577",
          "game_id": "sk48",
          "game_version": "d8078629",
          "published_at": "2026-09-16T17:05:07.924996Z",
          "score": 0,
          "model": "gpt-6-luna",
          "runner": "benchmark_agent",
          "config": "openai-gpt-6-luna-xhigh",
          "verified_publish": true,
          "levels_completed": 0,
          "actions": 305,
          "resets": 1,
          "state": "NOT_FINISHED",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            0,
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            305,
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            61,
            177,
            101,
            103,
            230,
            181,
            125,
            92
          ]
        },
        {
          "session_id": "2e0181b8-1e00-40de-a98e-e83134008fb5",
          "game_id": "sk48",
          "game_version": "d8078629",
          "published_at": "2026-09-16T16:51:57.404568Z",
          "score": 0,
          "model": "gpt-6-luna",
          "runner": "benchmark_agent",
          "config": "openai-gpt-6-luna-high",
          "verified_publish": true,
          "levels_completed": 0,
          "actions": 305,
          "resets": 1,
          "state": "GAME_OVER",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            0,
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            305,
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            61,
            177,
            101,
            103,
            230,
            181,
            125,
            92
          ]
        },
        {
          "session_id": "9cb8bfb1-73b3-4a4b-8dae-e10b52083909",
          "game_id": "sk48",
          "game_version": "d8078629",
          "published_at": "2026-09-16T16:19:30.457396Z",
          "score": 0,
          "model": "gpt-6-luna",
          "runner": "benchmark_agent",
          "config": "openai-gpt-6-luna-high-provider-adapter",
          "verified_publish": true,
          "levels_completed": 0,
          "actions": 305,
          "resets": 1,
          "state": "NOT_FINISHED",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            0,
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            305,
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            61,
            177,
            101,
            103,
            230,
            181,
            125,
            92
          ]
        },
        {
          "session_id": "8470d856-7576-41a9-b9a5-fcae66b4b2d4",
          "game_id": "sk48",
          "game_version": "d8078629",
          "published_at": "2026-09-16T15:29:28.714313Z",
          "score": 0,
          "model": "gpt-6-luna",
          "runner": "benchmark_agent",
          "config": "openai-gpt-6-luna-medium",
          "verified_publish": true,
          "levels_completed": 0,
          "actions": 305,
          "resets": 2,
          "state": "GAME_OVER",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            0,
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            305,
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            61,
            177,
            101,
            103,
            230,
            181,
            125,
            92
          ]
        },
        {
          "session_id": "9fd68fca-3984-4d97-937d-83e7cdad4d94",
          "game_id": "sk48",
          "game_version": "d8078629",
          "published_at": "2026-09-16T14:59:51.160053Z",
          "score": 18.89937996076685,
          "model": "gpt-6-1-sol",
          "runner": "benchmark_agent",
          "config": "openai-gpt-6-1-sol-max",
          "verified_publish": true,
          "levels_completed": 4,
          "actions": 877,
          "resets": 8,
          "state": "NOT_FINISHED",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            115,
            72.41355399408283,
            115,
            18.887642649860243,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            14,
            208,
            80,
            237,
            338,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            61,
            177,
            101,
            103,
            230,
            181,
            125,
            92
          ]
        },
        {
          "session_id": "5cb9df32-ba16-4d76-9256-40b499fe3ca4",
          "game_id": "sk48",
          "game_version": "d8078629",
          "published_at": "2026-09-16T14:44:25.623216Z",
          "score": 2.7777777777777777,
          "model": "gpt-6-1-sol",
          "runner": "benchmark_agent",
          "config": "openai-gpt-6-1-sol-low",
          "verified_publish": true,
          "levels_completed": 1,
          "actions": 48,
          "resets": 0,
          "state": "NOT_FINISHED",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            115,
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            16,
            32,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            61,
            177,
            101,
            103,
            230,
            181,
            125,
            92
          ]
        },
        {
          "session_id": "64e603f9-b5d1-496a-a4c2-1b3968eda8ef",
          "game_id": "sk48",
          "game_version": "d8078629",
          "published_at": "2026-09-16T11:59:15.838962Z",
          "score": 0,
          "model": "gpt-6-luna",
          "runner": "benchmark_agent",
          "config": "openai-gpt-6-luna-max",
          "verified_publish": true,
          "levels_completed": 0,
          "actions": 305,
          "resets": 3,
          "state": "NOT_FINISHED",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            0,
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            305,
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            61,
            177,
            101,
            103,
            230,
            181,
            125,
            92
          ]
        },
        {
          "session_id": "6935436d-18f8-4eb7-86d2-bec0e1e642e3",
          "game_id": "sk48",
          "game_version": "d8078629",
          "published_at": "2026-09-16T09:27:15.061498Z",
          "score": 0,
          "model": "gpt-6-luna",
          "runner": "benchmark_agent",
          "config": "openai-gpt-6-luna-medium-provider-adapter",
          "verified_publish": true,
          "levels_completed": 0,
          "actions": 305,
          "resets": 1,
          "state": "NOT_FINISHED",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            0,
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            305,
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            61,
            177,
            101,
            103,
            230,
            181,
            125,
            92
          ]
        },
        {
          "session_id": "fff4e810-bfc5-486e-af16-31118e241a04",
          "game_id": "sk48",
          "game_version": "d8078629",
          "published_at": "2026-09-16T08:59:43.727222Z",
          "score": 6.937458399431555,
          "model": "gpt-6-1-sol",
          "runner": "benchmark_agent",
          "config": "openai-gpt-6-1-sol-xhigh",
          "verified_publish": true,
          "levels_completed": 3,
          "actions": 1057,
          "resets": 14,
          "state": "NOT_FINISHED",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            115,
            29.120501189768,
            25.5025,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            14,
            328,
            200,
            515,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            61,
            177,
            101,
            103,
            230,
            181,
            125,
            92
          ]
        },
        {
          "session_id": "00039a91-362b-4844-a7f1-2ccd61866e42",
          "game_id": "sk48",
          "game_version": "d8078629",
          "published_at": "2026-09-16T08:36:20.001365Z",
          "score": 0,
          "model": "gpt-6-luna",
          "runner": "benchmark_agent",
          "config": "openai-gpt-6-luna-max-provider-adapter",
          "verified_publish": true,
          "levels_completed": 0,
          "actions": 305,
          "resets": 3,
          "state": "NOT_FINISHED",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            0,
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            305,
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            61,
            177,
            101,
            103,
            230,
            181,
            125,
            92
          ]
        },
        {
          "session_id": "663673c6-9d3d-4bd2-ad5c-ec575d1da738",
          "game_id": "sk48",
          "game_version": "d8078629",
          "published_at": "2026-09-16T08:33:34.516792Z",
          "score": 0,
          "model": "gpt-6-luna",
          "runner": "benchmark_agent",
          "config": "openai-gpt-6-luna-xhigh-provider-adapter",
          "verified_publish": true,
          "levels_completed": 0,
          "actions": 305,
          "resets": 1,
          "state": "GAME_OVER",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            0,
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            305,
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            61,
            177,
            101,
            103,
            230,
            181,
            125,
            92
          ]
        },
        {
          "session_id": "c40804c4-6ee2-4bcc-a794-be6bae2b4144",
          "game_id": "sk48",
          "game_version": "d8078629",
          "published_at": "2026-09-16T06:28:30.877437Z",
          "score": 1.9575850714347238,
          "model": "gpt-6-1-sol",
          "runner": "benchmark_agent",
          "config": "openai-gpt-6-1-sol-medium",
          "verified_publish": true,
          "levels_completed": 2,
          "actions": 1091,
          "resets": 13,
          "state": "NOT_FINISHED",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            4.849535377758083,
            32.81176359694599,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            277,
            309,
            505,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            61,
            177,
            101,
            103,
            230,
            181,
            125,
            92
          ]
        },
        {
          "session_id": "915e4646-9401-43ae-a00d-c11a8b840b16",
          "game_id": "sk48",
          "game_version": "d8078629",
          "published_at": "2026-09-16T06:19:41.917952Z",
          "score": 100,
          "model": "gpt-6-1-sol",
          "runner": "benchmark_agent",
          "config": "openai-gpt-6-1-sol-high-provider-adapter",
          "verified_publish": true,
          "levels_completed": 8,
          "actions": 615,
          "resets": 4,
          "state": "WIN",
          "completed": true,
          "number_of_levels": 0,
          "level_scores": [
            115,
            115,
            115,
            66.8241370622323,
            115,
            115,
            115,
            115
          ],
          "level_actions": [
            14,
            33,
            49,
            126,
            189,
            83,
            36,
            85
          ],
          "level_baseline_actions": [
            61,
            177,
            101,
            103,
            230,
            181,
            125,
            92
          ]
        },
        {
          "session_id": "6277fbd6-e9d7-46c8-abf9-602d91da6831",
          "game_id": "sk48",
          "game_version": "d8078629",
          "published_at": "2026-09-16T04:46:39.812971Z",
          "score": 100,
          "model": "gpt-6-1-sol",
          "runner": "benchmark_agent",
          "config": "openai-gpt-6-1-sol-xhigh-provider-adapter",
          "verified_publish": true,
          "levels_completed": 8,
          "actions": 427,
          "resets": 2,
          "state": "WIN",
          "completed": true,
          "number_of_levels": 0,
          "level_scores": [
            115,
            115,
            115,
            115,
            115,
            115,
            115,
            115
          ],
          "level_actions": [
            14,
            45,
            39,
            49,
            135,
            50,
            42,
            53
          ],
          "level_baseline_actions": [
            61,
            177,
            101,
            103,
            230,
            181,
            125,
            92
          ]
        },
        {
          "session_id": "c299de73-6d1c-45f5-be74-c3552c90af73",
          "game_id": "sk48",
          "game_version": "d8078629",
          "published_at": "2026-09-16T03:41:44.368972Z",
          "score": 100,
          "model": "gpt-6-1-sol",
          "runner": "benchmark_agent",
          "config": "openai-gpt-6-1-sol-max-provider-adapter",
          "verified_publish": true,
          "levels_completed": 8,
          "actions": 463,
          "resets": 4,
          "state": "WIN",
          "completed": true,
          "number_of_levels": 0,
          "level_scores": [
            115,
            115,
            115,
            115,
            115,
            115,
            115,
            115
          ],
          "level_actions": [
            14,
            77,
            36,
            53,
            101,
            53,
            64,
            65
          ],
          "level_baseline_actions": [
            61,
            177,
            101,
            103,
            230,
            181,
            125,
            92
          ]
        },
        {
          "session_id": "942249e7-ee25-4548-b55e-b12ac2aec082",
          "game_id": "sk48",
          "game_version": "d8078629",
          "published_at": "2026-09-12T22:45:56.763098Z",
          "score": 2.7777777777777777,
          "model": "gemini-3.8-flash",
          "runner": "benchmark_agent",
          "config": "google-gemini-3-8-flash-high",
          "verified_publish": true,
          "levels_completed": 1,
          "actions": 549,
          "resets": 3,
          "state": "GAME_OVER",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            115,
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            34,
            515,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            61,
            177,
            101,
            103,
            230,
            181,
            125,
            92
          ]
        },
        {
          "session_id": "f37d7da7-ef5d-4edf-ad2f-3dde5b3bed74",
          "game_id": "sk48",
          "game_version": "d8078629",
          "published_at": "2026-09-12T13:10:26.991729Z",
          "score": 2.7777777777777777,
          "model": "gemini-3.8-flash",
          "runner": "benchmark_agent",
          "config": "google-gemini-3-8-flash-low",
          "verified_publish": true,
          "levels_completed": 1,
          "actions": 911,
          "resets": 9,
          "state": "NOT_FINISHED",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            115,
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            26,
            885,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            61,
            177,
            101,
            103,
            230,
            181,
            125,
            92
          ]
        },
        {
          "session_id": "a4a7220b-f8a8-4393-bcc4-eb84be50fc66",
          "game_id": "sk48",
          "game_version": "d8078629",
          "published_at": "2026-09-09T19:46:05.173783Z",
          "score": 2.7777777777777777,
          "model": "gemini-3.8-flash",
          "runner": "benchmark_agent",
          "config": "google-gemini-3-8-flash-medium",
          "verified_publish": true,
          "levels_completed": 1,
          "actions": 561,
          "resets": 3,
          "state": "GAME_OVER",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            115,
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            30,
            531,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            61,
            177,
            101,
            103,
            230,
            181,
            125,
            92
          ]
        },
        {
          "session_id": "839a6187-dfe3-48a7-8d8a-6e32cb656200",
          "game_id": "sk48",
          "game_version": "d8078629",
          "published_at": "2026-08-11T10:49:39.241967Z",
          "score": 0,
          "model": "grok-4.6",
          "runner": "benchmark_agent",
          "config": "xai-grok-4-6-xhigh",
          "verified_publish": true,
          "levels_completed": 0,
          "actions": 305,
          "resets": 1,
          "state": "NOT_FINISHED",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            0,
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            305,
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            61,
            177,
            101,
            103,
            230,
            181,
            125,
            92
          ]
        },
        {
          "session_id": "c6f88b81-e660-426a-b632-c1b5b0fc966c",
          "game_id": "sk48",
          "game_version": "d8078629",
          "published_at": "2026-07-21T08:00:53.807092Z",
          "score": 8.333333333333332,
          "model": "claude-opus-5",
          "runner": "benchmark_agent",
          "config": "anthropic-claude-opus-5-high",
          "verified_publish": true,
          "levels_completed": 2,
          "actions": 684,
          "resets": 99,
          "state": "NOT_FINISHED",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            115,
            115,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            17,
            162,
            505,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            61,
            177,
            101,
            103,
            230,
            181,
            125,
            92
          ]
        },
        {
          "session_id": "bbbf1240-c6fd-41a2-9f4a-6d79687bbe3f",
          "game_id": "sk48",
          "game_version": "d8078629",
          "published_at": "2026-06-24T11:07:32.721822Z",
          "score": 0,
          "model": "gpt-5-6-luna",
          "runner": "benchmark_agent",
          "config": "openai-gpt-5-6-luna-low",
          "verified_publish": true,
          "levels_completed": 0,
          "actions": 305,
          "resets": 5,
          "state": "NOT_FINISHED",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            0,
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            305,
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            61,
            177,
            101,
            103,
            230,
            181,
            125,
            92
          ]
        },
        {
          "session_id": "6b3e0dc8-435b-4d47-b837-ec4dcd2fbc9d",
          "game_id": "sk48",
          "game_version": "d8078629",
          "published_at": "2026-06-24T10:43:33.391232Z",
          "score": 0,
          "model": "gpt-5-6-luna",
          "runner": "benchmark_agent",
          "config": "openai-gpt-5-6-luna-medium",
          "verified_publish": true,
          "levels_completed": 0,
          "actions": 305,
          "resets": 2,
          "state": "NOT_FINISHED",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            0,
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            305,
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            61,
            177,
            101,
            103,
            230,
            181,
            125,
            92
          ]
        },
        {
          "session_id": "8b1fc29d-dc7b-42e0-a91b-b9517e224430",
          "game_id": "sk48",
          "game_version": "d8078629",
          "published_at": "2026-06-24T03:56:15.285437Z",
          "score": 0,
          "model": "gpt-5-6-luna",
          "runner": "benchmark_agent",
          "config": "openai-gpt-5-6-luna-high",
          "verified_publish": true,
          "levels_completed": 0,
          "actions": 305,
          "resets": 38,
          "state": "NOT_FINISHED",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            0,
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            305,
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            61,
            177,
            101,
            103,
            230,
            181,
            125,
            92
          ]
        },
        {
          "session_id": "a4d9d28e-0547-4e1e-a671-3724f6c73896",
          "game_id": "sk48",
          "game_version": "d8078629",
          "published_at": "2026-06-24T02:30:31.633712Z",
          "score": 0,
          "model": "gpt-5-6-luna",
          "runner": "benchmark_agent",
          "config": "openai-gpt-5-6-luna-xhigh",
          "verified_publish": true,
          "levels_completed": 0,
          "actions": 305,
          "resets": 43,
          "state": "NOT_FINISHED",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            0,
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            305,
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            61,
            177,
            101,
            103,
            230,
            181,
            125,
            92
          ]
        },
        {
          "session_id": "8572f6cd-9ba3-4cb5-80ed-69654102bd76",
          "game_id": "sk48",
          "game_version": "d8078629",
          "published_at": "2026-06-23T04:35:03.824187Z",
          "score": 0,
          "model": "gpt-5-6-luna",
          "runner": "benchmark_agent",
          "config": "openai-gpt-5-6-luna-max",
          "verified_publish": true,
          "levels_completed": 0,
          "actions": 305,
          "resets": 23,
          "state": "NOT_FINISHED",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            0,
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            305,
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            61,
            177,
            101,
            103,
            230,
            181,
            125,
            92
          ]
        },
        {
          "session_id": "3598d8da-209e-43d7-ae0b-93da70e60bb3",
          "game_id": "sk48",
          "game_version": "d8078629",
          "published_at": "2026-06-20T12:39:32.903369Z",
          "score": 0.2389078936554898,
          "model": "gpt-5-6-terra",
          "runner": "benchmark_agent",
          "config": "openai-gpt-5-6-terra-xhigh",
          "verified_publish": true,
          "levels_completed": 1,
          "actions": 1093,
          "resets": 15,
          "state": "NOT_FINISHED",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            8.600684171597633,
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            208,
            885,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            61,
            177,
            101,
            103,
            230,
            181,
            125,
            92
          ]
        },
        {
          "session_id": "b60a3b06-d7ab-419f-90ae-3b0159eafc7e",
          "game_id": "sk48",
          "game_version": "d8078629",
          "published_at": "2026-06-20T12:05:58.278407Z",
          "score": 0,
          "model": "gpt-5-6-terra",
          "runner": "benchmark_agent",
          "config": "openai-gpt-5-6-terra-low",
          "verified_publish": true,
          "levels_completed": 0,
          "actions": 305,
          "resets": 2,
          "state": "GAME_OVER",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            0,
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            305,
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            61,
            177,
            101,
            103,
            230,
            181,
            125,
            92
          ]
        },
        {
          "session_id": "ef23d316-5e46-4711-ba0a-8bfe7f2168a9",
          "game_id": "sk48",
          "game_version": "d8078629",
          "published_at": "2026-06-20T09:23:40.779717Z",
          "score": 0,
          "model": "gpt-5-6-terra",
          "runner": "benchmark_agent",
          "config": "openai-gpt-5-6-terra-high",
          "verified_publish": true,
          "levels_completed": 0,
          "actions": 305,
          "resets": 1,
          "state": "NOT_FINISHED",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            0,
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            305,
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            61,
            177,
            101,
            103,
            230,
            181,
            125,
            92
          ]
        },
        {
          "session_id": "f0563c6f-ade5-45f0-87ff-07d2aa5fc08b",
          "game_id": "sk48",
          "game_version": "d8078629",
          "published_at": "2026-06-20T09:18:23.722104Z",
          "score": 0,
          "model": "gpt-5-6-terra",
          "runner": "benchmark_agent",
          "config": "openai-gpt-5-6-terra-medium",
          "verified_publish": true,
          "levels_completed": 0,
          "actions": 305,
          "resets": 1,
          "state": "GAME_OVER",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            0,
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            305,
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            61,
            177,
            101,
            103,
            230,
            181,
            125,
            92
          ]
        },
        {
          "session_id": "5f6d1fcb-c6f1-4a7b-90a0-2ba7855db1b9",
          "game_id": "sk48",
          "game_version": "d8078629",
          "published_at": "2026-06-19T02:00:32.308063Z",
          "score": 0,
          "model": "gpt-5-6-terra",
          "runner": "benchmark_agent",
          "config": "openai-gpt-5-6-terra-max",
          "verified_publish": true,
          "levels_completed": 0,
          "actions": 305,
          "resets": 14,
          "state": "NOT_FINISHED",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            0,
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            305,
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            61,
            177,
            101,
            103,
            230,
            181,
            125,
            92
          ]
        },
        {
          "session_id": "de4a5410-6835-41ff-a0f4-c0e45102efd3",
          "game_id": "sk48",
          "game_version": "d8078629",
          "published_at": "2026-06-16T07:37:00.3054Z",
          "score": 1.1950642977351267,
          "model": "gpt-5-6-sol",
          "runner": "benchmark_agent",
          "config": "openai-gpt-5-6-sol-xhigh",
          "verified_publish": true,
          "levels_completed": 1,
          "actions": 978,
          "resets": 12,
          "state": "NOT_FINISHED",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            43.02231471846456,
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            93,
            885,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            61,
            177,
            101,
            103,
            230,
            181,
            125,
            92
          ]
        },
        {
          "session_id": "7d69a3eb-f4c5-4add-8d67-5e8c0ecc6b60",
          "game_id": "sk48",
          "game_version": "d8078629",
          "published_at": "2026-06-13T04:42:13.349388Z",
          "score": 0,
          "model": "gpt-5-6-sol",
          "runner": "benchmark_agent",
          "config": "openai-gpt-5-6-sol-max",
          "verified_publish": true,
          "levels_completed": 0,
          "actions": 305,
          "resets": 3,
          "state": "NOT_FINISHED",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            0,
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            305,
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            61,
            177,
            101,
            103,
            230,
            181,
            125,
            92
          ]
        },
        {
          "session_id": "9fa6fcb2-f2d7-4f79-94ba-89c53b4dd5a3",
          "game_id": "sk48",
          "game_version": "d8078629",
          "published_at": "2026-06-10T17:41:00.328342Z",
          "score": 2.7777777777777777,
          "model": "gpt-5-6-sol",
          "runner": "benchmark_agent",
          "config": "openai-gpt-5-6-sol-high",
          "verified_publish": true,
          "levels_completed": 1,
          "actions": 919,
          "resets": 8,
          "state": "GAME_OVER",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            115,
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            34,
            885,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            61,
            177,
            101,
            103,
            230,
            181,
            125,
            92
          ]
        },
        {
          "session_id": "bad8a233-dc2f-449f-9b43-100d58456dcb",
          "game_id": "sk48",
          "game_version": "d8078629",
          "published_at": "2026-06-10T08:30:17.325111Z",
          "score": 0,
          "model": "gpt-5-6-sol",
          "runner": "benchmark_agent",
          "config": "openai-gpt-5-6-sol-medium",
          "verified_publish": true,
          "levels_completed": 0,
          "actions": 305,
          "resets": 5,
          "state": "NOT_FINISHED",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            0,
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            305,
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            61,
            177,
            101,
            103,
            230,
            181,
            125,
            92
          ]
        },
        {
          "session_id": "11fbbfc4-1d25-476e-9fc0-8dec427e2a2a",
          "game_id": "sk48",
          "game_version": "d8078629",
          "published_at": "2026-06-09T21:49:43.838365Z",
          "score": 0,
          "model": "gpt-5-6-sol",
          "runner": "benchmark_agent",
          "config": "openai-gpt-5-6-sol-low",
          "verified_publish": true,
          "levels_completed": 0,
          "actions": 305,
          "resets": 3,
          "state": "NOT_FINISHED",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            0,
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            305,
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            61,
            177,
            101,
            103,
            230,
            181,
            125,
            92
          ]
        },
        {
          "session_id": "ab291a16-ff6b-485c-98da-3e07c27b17de",
          "game_id": "sk48",
          "game_version": "d8078629",
          "published_at": "2026-05-31T14:56:34.992584Z",
          "score": 0,
          "model": "claude-opus-4-8",
          "runner": "benchmark_agent",
          "config": "anthropic-opus-4-8-high",
          "verified_publish": true,
          "levels_completed": 0,
          "actions": 305,
          "resets": 5,
          "state": "NOT_FINISHED",
          "completed": false,
          "number_of_levels": 8,
          "level_scores": [
            0,
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            305,
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            61,
            177,
            101,
            103,
            230,
            181,
            125,
            92
          ]
        },
        {
          "session_id": "1c3674d9-586c-4f4a-9f44-b51031070ed2",
          "game_id": "sk48",
          "game_version": "d8078629",
          "published_at": "2026-05-01T16:12:21.627016Z",
          "score": 0,
          "model": "claude-opus-4-7",
          "runner": "benchmark_agent",
          "config": "anthropic-opus-4-7-high",
          "verified_publish": true,
          "levels_completed": 0,
          "actions": 305,
          "resets": 5,
          "state": "NOT_FINISHED",
          "completed": false,
          "number_of_levels": 8,
          "level_scores": [
            0,
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            305,
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            61,
            177,
            101,
            103,
            230,
            181,
            125,
            92
          ]
        },
        {
          "session_id": "225cdfe3-a9db-4d77-9deb-2aa9289c3f16",
          "game_id": "sk48",
          "game_version": "d8078629",
          "published_at": "2026-04-30T23:21:21.713556Z",
          "score": 0,
          "model": "gpt-5.5-2026-04-23",
          "runner": "benchmark_agent",
          "config": "openai-gpt-5-5-2026-04-23-high",
          "verified_publish": true,
          "levels_completed": 0,
          "actions": 305,
          "resets": 2,
          "state": "GAME_OVER",
          "completed": false,
          "number_of_levels": 8,
          "level_scores": [
            0,
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            305,
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            61,
            177,
            101,
            103,
            230,
            181,
            125,
            92
          ]
        },
        {
          "session_id": "3d02671a-157d-4753-95c8-b4065e1a6c6e",
          "game_id": "sk48",
          "game_version": "d8078629",
          "published_at": "2026-04-14T17:43:55.054632Z",
          "score": 0,
          "model": "grok-4.20-beta-0309-reasoning",
          "runner": "benchmark_agent",
          "config": "xai-grok-4-20-beta-0309-reasoning",
          "verified_publish": true,
          "levels_completed": 0,
          "actions": 18,
          "resets": 0,
          "state": "NOT_FINISHED",
          "completed": false,
          "number_of_levels": 8,
          "level_scores": [
            0,
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            18,
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            61,
            177,
            101,
            103,
            230,
            181,
            125,
            92
          ]
        },
        {
          "session_id": "14ab5f55-7002-467f-ae9b-e7d1e387f256",
          "game_id": "sk48",
          "game_version": "d8078629",
          "published_at": "2026-04-14T17:43:54.732434Z",
          "score": 0,
          "model": "gpt-5.4-2026-03-05",
          "runner": "benchmark_agent",
          "config": "openai-gpt-5-4-2026-03-05-high",
          "verified_publish": true,
          "levels_completed": 0,
          "actions": 305,
          "resets": 1,
          "state": "GAME_OVER",
          "completed": false,
          "number_of_levels": 8,
          "level_scores": [
            0,
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            305,
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            61,
            177,
            101,
            103,
            230,
            181,
            125,
            92
          ]
        },
        {
          "session_id": "0099e80b-f345-48d6-a97b-45da2389749e",
          "game_id": "sk48",
          "game_version": "d8078629",
          "published_at": "2026-04-14T17:43:54.362929Z",
          "score": 0,
          "model": "gemini-3.1-pro-preview",
          "runner": "benchmark_agent",
          "config": "google-gemini-3-1-pro-preview",
          "verified_publish": true,
          "levels_completed": 0,
          "actions": 305,
          "resets": 3,
          "state": "NOT_FINISHED",
          "completed": false,
          "number_of_levels": 8,
          "level_scores": [
            0,
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            305,
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            61,
            177,
            101,
            103,
            230,
            181,
            125,
            92
          ]
        },
        {
          "session_id": "f6ec9d34-8830-41a1-be7d-3a159c88c2d3",
          "game_id": "sk48",
          "game_version": "d8078629",
          "published_at": "2026-04-14T17:43:54.028713Z",
          "score": 0,
          "model": "claude-opus-4-6",
          "runner": "benchmark_agent",
          "config": "anthropic-opus-4-6-max-effort",
          "verified_publish": true,
          "levels_completed": 0,
          "actions": 305,
          "resets": 25,
          "state": "NOT_FINISHED",
          "completed": false,
          "number_of_levels": 8,
          "level_scores": [
            0,
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            305,
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            61,
            177,
            101,
            103,
            230,
            181,
            125,
            92
          ]
        }
      ]
    },
    {
      "game": "s5i5",
      "source": "https://arcprize.org/tasks/s5i5",
      "runs": [
        {
          "session_id": "2b20cc81-21ba-4991-84c6-d184391b9fd3",
          "game_id": "s5i5",
          "game_version": "18d95033",
          "published_at": "2026-09-25T15:35:15.754694Z",
          "score": 1.671173009556848,
          "model": "gpt-6-sol",
          "runner": "benchmark_agent",
          "config": "openai-gpt-6-sol-xhigh",
          "verified_publish": true,
          "levels_completed": 2,
          "actions": 726,
          "resets": 3,
          "state": "GAME_OVER",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            19.753086419753085,
            20.20457096214672,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            45,
            198,
            483,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            20,
            89,
            106,
            54,
            162,
            38,
            86,
            83
          ]
        },
        {
          "session_id": "53d56d7b-a2a9-4e26-a501-dad474df1c5e",
          "game_id": "s5i5",
          "game_version": "18d95033",
          "published_at": "2026-09-25T15:24:23.400436Z",
          "score": 7.545091918140825,
          "model": "gpt-6-sol",
          "runner": "benchmark_agent",
          "config": "openai-gpt-6-sol-max",
          "verified_publish": true,
          "levels_completed": 2,
          "actions": 294,
          "resets": 2,
          "state": "GAME_OVER",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            41.62330905306972,
            115,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            31,
            41,
            222,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            20,
            89,
            106,
            54,
            162,
            38,
            86,
            83
          ]
        },
        {
          "session_id": "7b3e0d9e-c5ba-4003-ad95-f6b6d8f4e03c",
          "game_id": "s5i5",
          "game_version": "18d95033",
          "published_at": "2026-09-25T08:36:03.106995Z",
          "score": 15.740140762449771,
          "model": "gpt-6-sol",
          "runner": "benchmark_agent",
          "config": "openai-gpt-6-sol-high-provider-adapter",
          "verified_publish": true,
          "levels_completed": 4,
          "actions": 1200,
          "resets": 6,
          "state": "GAME_OVER",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            18.10774105930285,
            85.94835069444446,
            115,
            7.91015625,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            47,
            96,
            55,
            192,
            810,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            20,
            89,
            106,
            54,
            162,
            38,
            86,
            83
          ]
        },
        {
          "session_id": "672f7a0b-ea46-4bd6-888d-247e900d24d7",
          "game_id": "s5i5",
          "game_version": "18d95033",
          "published_at": "2026-09-25T06:29:27.994145Z",
          "score": 0,
          "model": "gpt-6-sol",
          "runner": "benchmark_agent",
          "config": "openai-gpt-6-sol-none-provider-adapter",
          "verified_publish": true,
          "levels_completed": 0,
          "actions": 100,
          "resets": 2,
          "state": "GAME_OVER",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            0,
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            100,
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            20,
            89,
            106,
            54,
            162,
            38,
            86,
            83
          ]
        },
        {
          "session_id": "68341e12-a50d-4caa-8c89-5f7b9d20d9e4",
          "game_id": "s5i5",
          "game_version": "18d95033",
          "published_at": "2026-09-25T06:19:08.537129Z",
          "score": 0,
          "model": "gpt-6-sol",
          "runner": "benchmark_agent",
          "config": "openai-gpt-6-sol-none",
          "verified_publish": true,
          "levels_completed": 0,
          "actions": 100,
          "resets": 1,
          "state": "GAME_OVER",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            0,
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            100,
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            20,
            89,
            106,
            54,
            162,
            38,
            86,
            83
          ]
        },
        {
          "session_id": "bc4beef5-e4ee-4a96-9896-37fd86990dc6",
          "game_id": "s5i5",
          "game_version": "18d95033",
          "published_at": "2026-09-25T06:05:23.751746Z",
          "score": 0,
          "model": "gpt-6-sol",
          "runner": "benchmark_agent",
          "config": "openai-gpt-6-sol-low-provider-adapter",
          "verified_publish": true,
          "levels_completed": 0,
          "actions": 100,
          "resets": 2,
          "state": "GAME_OVER",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            0,
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            100,
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            20,
            89,
            106,
            54,
            162,
            38,
            86,
            83
          ]
        },
        {
          "session_id": "3813d293-3d74-4f0e-af3d-2e8f5127b597",
          "game_id": "s5i5",
          "game_version": "18d95033",
          "published_at": "2026-09-25T05:53:23.763029Z",
          "score": 14.71426367028589,
          "model": "gpt-6-sol",
          "runner": "benchmark_agent",
          "config": "openai-gpt-6-sol-xhigh-provider-adapter",
          "verified_publish": true,
          "levels_completed": 3,
          "actions": 483,
          "resets": 3,
          "state": "NOT_FINISHED",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            16.000000000000004,
            115,
            94.5711640434307,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            50,
            54,
            109,
            270,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            20,
            89,
            106,
            54,
            162,
            38,
            86,
            83
          ]
        },
        {
          "session_id": "f4eae844-58a8-4314-94a4-e553986e70b1",
          "game_id": "s5i5",
          "game_version": "18d95033",
          "published_at": "2026-09-25T05:53:18.198261Z",
          "score": 0,
          "model": "gpt-6-sol",
          "runner": "benchmark_agent",
          "config": "openai-gpt-6-sol-low",
          "verified_publish": true,
          "levels_completed": 0,
          "actions": 100,
          "resets": 2,
          "state": "GAME_OVER",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            0,
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            100,
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            20,
            89,
            106,
            54,
            162,
            38,
            86,
            83
          ]
        },
        {
          "session_id": "59916881-b1f2-40ba-9e3e-d4418879c7ff",
          "game_id": "s5i5",
          "game_version": "18d95033",
          "published_at": "2026-09-25T05:35:24.213713Z",
          "score": 0,
          "model": "gpt-6-sol",
          "runner": "benchmark_agent",
          "config": "openai-gpt-6-sol-medium-provider-adapter",
          "verified_publish": true,
          "levels_completed": 0,
          "actions": 100,
          "resets": 2,
          "state": "GAME_OVER",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            0,
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            100,
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            20,
            89,
            106,
            54,
            162,
            38,
            86,
            83
          ]
        },
        {
          "session_id": "e90e28ec-6fc7-473f-a2fb-b177e3ebfdcc",
          "game_id": "s5i5",
          "game_version": "18d95033",
          "published_at": "2026-09-25T05:13:23.42073Z",
          "score": 0,
          "model": "gpt-6-sol",
          "runner": "benchmark_agent",
          "config": "openai-gpt-6-sol-medium",
          "verified_publish": true,
          "levels_completed": 0,
          "actions": 100,
          "resets": 2,
          "state": "GAME_OVER",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            0,
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            100,
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            20,
            89,
            106,
            54,
            162,
            38,
            86,
            83
          ]
        },
        {
          "session_id": "a4eed85e-e606-4450-9776-1500cd1a3623",
          "game_id": "s5i5",
          "game_version": "18d95033",
          "published_at": "2026-09-25T04:35:35.546995Z",
          "score": 0,
          "model": "gpt-6-sol",
          "runner": "benchmark_agent",
          "config": "openai-gpt-6-sol-high",
          "verified_publish": true,
          "levels_completed": 0,
          "actions": 100,
          "resets": 1,
          "state": "GAME_OVER",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            0,
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            100,
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            20,
            89,
            106,
            54,
            162,
            38,
            86,
            83
          ]
        },
        {
          "session_id": "e2587d60-72c2-4301-987c-95b1a7ab1bf8",
          "game_id": "s5i5",
          "game_version": "18d95033",
          "published_at": "2026-09-21T14:33:27.358498Z",
          "score": 77.83216601053385,
          "model": "gemini-3.8-flash",
          "runner": "benchmark_agent",
          "config": "google-gemini-3-8-flash-low-provider-adapter",
          "verified_publish": true,
          "levels_completed": 8,
          "actions": 928,
          "resets": 4,
          "state": "WIN",
          "completed": true,
          "number_of_levels": 0,
          "level_scores": [
            100,
            115,
            115,
            115,
            11.020223058317654,
            11.11111111111111,
            89.31288491728053,
            115
          ],
          "level_actions": [
            20,
            28,
            83,
            49,
            488,
            114,
            91,
            55
          ],
          "level_baseline_actions": [
            20,
            89,
            106,
            54,
            162,
            38,
            86,
            83
          ]
        },
        {
          "session_id": "114e5ba9-3ea1-447a-9e95-bc042f6c5e86",
          "game_id": "s5i5",
          "game_version": "18d95033",
          "published_at": "2026-09-20T09:40:14.361794Z",
          "score": 67.48885545286883,
          "model": "gemini-3.8-flash",
          "runner": "benchmark_agent",
          "config": "google-gemini-3-8-flash-medium-provider-adapter",
          "verified_publish": true,
          "levels_completed": 8,
          "actions": 761,
          "resets": 2,
          "state": "WIN",
          "completed": true,
          "number_of_levels": 0,
          "level_scores": [
            90.70294784580499,
            115,
            115,
            4.8579758433985845,
            115,
            27.097016325764685,
            44.44444444444444,
            86.97134200227244
          ],
          "level_actions": [
            21,
            26,
            80,
            245,
            98,
            73,
            129,
            89
          ],
          "level_baseline_actions": [
            20,
            89,
            106,
            54,
            162,
            38,
            86,
            83
          ]
        },
        {
          "session_id": "8909dd03-0b21-42bb-9b42-4c79a64b2c58",
          "game_id": "s5i5",
          "game_version": "18d95033",
          "published_at": "2026-09-17T20:28:34.224405Z",
          "score": 100,
          "model": "gemini-3.8-flash",
          "runner": "benchmark_agent",
          "config": "google-gemini-3-8-flash-high-provider-adapter",
          "verified_publish": true,
          "levels_completed": 8,
          "actions": 330,
          "resets": 0,
          "state": "WIN",
          "completed": true,
          "number_of_levels": 0,
          "level_scores": [
            115,
            115,
            115,
            115,
            115,
            115,
            72.50269581413586,
            115
          ],
          "level_actions": [
            16,
            26,
            45,
            32,
            34,
            29,
            101,
            47
          ],
          "level_baseline_actions": [
            20,
            89,
            106,
            54,
            162,
            38,
            86,
            83
          ]
        },
        {
          "session_id": "9f91b5b5-1612-4395-9b5b-ce2497a99e8f",
          "game_id": "s5i5",
          "game_version": "18d95033",
          "published_at": "2026-09-16T19:25:26.406675Z",
          "score": 0,
          "model": "gpt-6-luna",
          "runner": "benchmark_agent",
          "config": "openai-gpt-6-luna-none-provider-adapter",
          "verified_publish": true,
          "levels_completed": 0,
          "actions": 100,
          "resets": 4,
          "state": "GAME_OVER",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            0,
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            100,
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            20,
            89,
            106,
            54,
            162,
            38,
            86,
            83
          ]
        },
        {
          "session_id": "ffe4bd65-d478-453b-acf4-1ab62af4bc7b",
          "game_id": "s5i5",
          "game_version": "18d95033",
          "published_at": "2026-09-16T19:22:38.860893Z",
          "score": 0,
          "model": "gpt-6-luna",
          "runner": "benchmark_agent",
          "config": "openai-gpt-6-luna-none",
          "verified_publish": true,
          "levels_completed": 0,
          "actions": 100,
          "resets": 2,
          "state": "NOT_FINISHED",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            0,
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            100,
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            20,
            89,
            106,
            54,
            162,
            38,
            86,
            83
          ]
        },
        {
          "session_id": "32a4d4f1-6409-4122-a968-b07eff97d1b6",
          "game_id": "s5i5",
          "game_version": "18d95033",
          "published_at": "2026-09-16T19:22:10.760208Z",
          "score": 0,
          "model": "gpt-6-luna",
          "runner": "benchmark_agent",
          "config": "openai-gpt-6-luna-low-provider-adapter",
          "verified_publish": true,
          "levels_completed": 0,
          "actions": 100,
          "resets": 5,
          "state": "GAME_OVER",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            0,
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            100,
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            20,
            89,
            106,
            54,
            162,
            38,
            86,
            83
          ]
        },
        {
          "session_id": "f388be28-65e1-44f9-9798-d4c84e381aac",
          "game_id": "s5i5",
          "game_version": "18d95033",
          "published_at": "2026-09-16T16:15:40.971048Z",
          "score": 0,
          "model": "gpt-6-luna",
          "runner": "benchmark_agent",
          "config": "openai-gpt-6-luna-xhigh",
          "verified_publish": true,
          "levels_completed": 0,
          "actions": 100,
          "resets": 4,
          "state": "GAME_OVER",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            0,
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            100,
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            20,
            89,
            106,
            54,
            162,
            38,
            86,
            83
          ]
        },
        {
          "session_id": "665747a9-402f-441d-88a7-ff8fa472baa2",
          "game_id": "s5i5",
          "game_version": "18d95033",
          "published_at": "2026-09-16T09:09:22.58459Z",
          "score": 0,
          "model": "gpt-6-luna",
          "runner": "benchmark_agent",
          "config": "openai-gpt-6-luna-low",
          "verified_publish": true,
          "levels_completed": 0,
          "actions": 100,
          "resets": 2,
          "state": "NOT_FINISHED",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            0,
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            100,
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            20,
            89,
            106,
            54,
            162,
            38,
            86,
            83
          ]
        },
        {
          "session_id": "568287bb-5bc8-4d06-91ed-df447c0253a4",
          "game_id": "s5i5",
          "game_version": "18d95033",
          "published_at": "2026-09-16T09:00:35.455094Z",
          "score": 0,
          "model": "gpt-6-luna",
          "runner": "benchmark_agent",
          "config": "openai-gpt-6-luna-medium-provider-adapter",
          "verified_publish": true,
          "levels_completed": 0,
          "actions": 100,
          "resets": 4,
          "state": "GAME_OVER",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            0,
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            100,
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            20,
            89,
            106,
            54,
            162,
            38,
            86,
            83
          ]
        },
        {
          "session_id": "404e1962-fc82-47d7-9dcb-2a68e5ebfe85",
          "game_id": "s5i5",
          "game_version": "18d95033",
          "published_at": "2026-09-16T08:53:36.083094Z",
          "score": 0,
          "model": "gpt-6-luna",
          "runner": "benchmark_agent",
          "config": "openai-gpt-6-luna-medium",
          "verified_publish": true,
          "levels_completed": 0,
          "actions": 100,
          "resets": 2,
          "state": "NOT_FINISHED",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            0,
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            100,
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            20,
            89,
            106,
            54,
            162,
            38,
            86,
            83
          ]
        },
        {
          "session_id": "08559d56-9412-4212-88f0-444c8428e9f6",
          "game_id": "s5i5",
          "game_version": "18d95033",
          "published_at": "2026-09-16T08:38:20.116217Z",
          "score": 0,
          "model": "gpt-6-luna",
          "runner": "benchmark_agent",
          "config": "openai-gpt-6-luna-high",
          "verified_publish": true,
          "levels_completed": 0,
          "actions": 100,
          "resets": 3,
          "state": "GAME_OVER",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            0,
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            100,
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            20,
            89,
            106,
            54,
            162,
            38,
            86,
            83
          ]
        },
        {
          "session_id": "5f509ff1-f21a-4bd9-9638-ff5cb4f28cfb",
          "game_id": "s5i5",
          "game_version": "18d95033",
          "published_at": "2026-09-16T08:34:32.142258Z",
          "score": 0,
          "model": "gpt-6-luna",
          "runner": "benchmark_agent",
          "config": "openai-gpt-6-luna-high-provider-adapter",
          "verified_publish": true,
          "levels_completed": 0,
          "actions": 100,
          "resets": 5,
          "state": "NOT_FINISHED",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            0,
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            100,
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            20,
            89,
            106,
            54,
            162,
            38,
            86,
            83
          ]
        },
        {
          "session_id": "0de0b37e-11ad-4ca7-a7a6-63d47cf056a3",
          "game_id": "s5i5",
          "game_version": "18d95033",
          "published_at": "2026-09-16T07:34:38.895474Z",
          "score": 0,
          "model": "gpt-6-luna",
          "runner": "benchmark_agent",
          "config": "openai-gpt-6-luna-xhigh-provider-adapter",
          "verified_publish": true,
          "levels_completed": 0,
          "actions": 100,
          "resets": 5,
          "state": "GAME_OVER",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            0,
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            100,
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            20,
            89,
            106,
            54,
            162,
            38,
            86,
            83
          ]
        },
        {
          "session_id": "c270b511-ed68-4195-9e25-1a4070206dca",
          "game_id": "s5i5",
          "game_version": "18d95033",
          "published_at": "2026-09-16T07:01:59.107919Z",
          "score": 0,
          "model": "gpt-6-luna",
          "runner": "benchmark_agent",
          "config": "openai-gpt-6-luna-max-provider-adapter",
          "verified_publish": true,
          "levels_completed": 0,
          "actions": 100,
          "resets": 5,
          "state": "NOT_FINISHED",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            0,
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            100,
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            20,
            89,
            106,
            54,
            162,
            38,
            86,
            83
          ]
        },
        {
          "session_id": "99fd8726-523e-4197-967c-4cc2051d22bc",
          "game_id": "s5i5",
          "game_version": "18d95033",
          "published_at": "2026-09-16T06:47:38.883601Z",
          "score": 16.666666666666664,
          "model": "gpt-6-1-sol",
          "runner": "benchmark_agent",
          "config": "openai-gpt-6-1-sol-low",
          "verified_publish": true,
          "levels_completed": 3,
          "actions": 394,
          "resets": 2,
          "state": "GAME_OVER",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            36.73094582185492,
            115,
            115,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            33,
            30,
            61,
            270,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            20,
            89,
            106,
            54,
            162,
            38,
            86,
            83
          ]
        },
        {
          "session_id": "b1e46794-dc96-4c81-98c2-038eb74bd758",
          "game_id": "s5i5",
          "game_version": "18d95033",
          "published_at": "2026-09-16T05:53:01.091426Z",
          "score": 97.0285934495284,
          "model": "gpt-6-1-sol",
          "runner": "benchmark_agent",
          "config": "openai-gpt-6-1-sol-medium",
          "verified_publish": true,
          "levels_completed": 8,
          "actions": 489,
          "resets": 0,
          "state": "WIN",
          "completed": true,
          "number_of_levels": 0,
          "level_scores": [
            115,
            115,
            115,
            115,
            115,
            115,
            22.5756234547175,
            115
          ],
          "level_actions": [
            17,
            32,
            42,
            44,
            75,
            29,
            181,
            69
          ],
          "level_baseline_actions": [
            20,
            89,
            106,
            54,
            162,
            38,
            86,
            83
          ]
        },
        {
          "session_id": "c4cf3b9c-e3ae-4840-9f97-e6d98eb117a6",
          "game_id": "s5i5",
          "game_version": "18d95033",
          "published_at": "2026-09-16T04:40:45.444469Z",
          "score": 100,
          "model": "gpt-6-1-sol",
          "runner": "benchmark_agent",
          "config": "openai-gpt-6-1-sol-high-provider-adapter",
          "verified_publish": true,
          "levels_completed": 8,
          "actions": 266,
          "resets": 0,
          "state": "WIN",
          "completed": true,
          "number_of_levels": 0,
          "level_scores": [
            115,
            115,
            115,
            115,
            115,
            115,
            115,
            115
          ],
          "level_actions": [
            13,
            26,
            40,
            31,
            40,
            29,
            46,
            41
          ],
          "level_baseline_actions": [
            20,
            89,
            106,
            54,
            162,
            38,
            86,
            83
          ]
        },
        {
          "session_id": "ebc5a49d-c102-44a4-89db-3a00415cf389",
          "game_id": "s5i5",
          "game_version": "18d95033",
          "published_at": "2026-09-16T04:30:53.86302Z",
          "score": 100,
          "model": "gpt-6-1-sol",
          "runner": "benchmark_agent",
          "config": "openai-gpt-6-1-sol-high",
          "verified_publish": true,
          "levels_completed": 8,
          "actions": 297,
          "resets": 2,
          "state": "WIN",
          "completed": true,
          "number_of_levels": 0,
          "level_scores": [
            115,
            115,
            115,
            115,
            115,
            115,
            115,
            115
          ],
          "level_actions": [
            15,
            26,
            38,
            48,
            30,
            29,
            66,
            45
          ],
          "level_baseline_actions": [
            20,
            89,
            106,
            54,
            162,
            38,
            86,
            83
          ]
        },
        {
          "session_id": "fb1de981-c8c0-43ad-9821-ad95e9deb84a",
          "game_id": "s5i5",
          "game_version": "18d95033",
          "published_at": "2026-09-16T03:51:13.996Z",
          "score": 0,
          "model": "gpt-6-luna",
          "runner": "benchmark_agent",
          "config": "openai-gpt-6-luna-max",
          "verified_publish": true,
          "levels_completed": 0,
          "actions": 100,
          "resets": 6,
          "state": "GAME_OVER",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            0,
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            100,
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            20,
            89,
            106,
            54,
            162,
            38,
            86,
            83
          ]
        },
        {
          "session_id": "3bf007fe-8c03-486d-84df-9973305d4518",
          "game_id": "s5i5",
          "game_version": "18d95033",
          "published_at": "2026-09-16T03:36:19.799244Z",
          "score": 100,
          "model": "gpt-6-1-sol",
          "runner": "benchmark_agent",
          "config": "openai-gpt-6-1-sol-xhigh",
          "verified_publish": true,
          "levels_completed": 8,
          "actions": 274,
          "resets": 0,
          "state": "WIN",
          "completed": true,
          "number_of_levels": 0,
          "level_scores": [
            115,
            115,
            115,
            115,
            115,
            115,
            115,
            115
          ],
          "level_actions": [
            18,
            26,
            40,
            30,
            35,
            27,
            51,
            47
          ],
          "level_baseline_actions": [
            20,
            89,
            106,
            54,
            162,
            38,
            86,
            83
          ]
        },
        {
          "session_id": "64e85ace-909a-418a-b84a-98869ab2f5f3",
          "game_id": "s5i5",
          "game_version": "18d95033",
          "published_at": "2026-09-16T03:18:47.640274Z",
          "score": 100,
          "model": "gpt-6-1-sol",
          "runner": "benchmark_agent",
          "config": "openai-gpt-6-1-sol-xhigh-provider-adapter",
          "verified_publish": true,
          "levels_completed": 8,
          "actions": 277,
          "resets": 0,
          "state": "WIN",
          "completed": true,
          "number_of_levels": 0,
          "level_scores": [
            115,
            115,
            115,
            115,
            115,
            115,
            115,
            115
          ],
          "level_actions": [
            18,
            26,
            39,
            34,
            31,
            27,
            58,
            44
          ],
          "level_baseline_actions": [
            20,
            89,
            106,
            54,
            162,
            38,
            86,
            83
          ]
        },
        {
          "session_id": "568ece0a-c637-4ffd-8f37-82a335576069",
          "game_id": "s5i5",
          "game_version": "18d95033",
          "published_at": "2026-09-16T01:38:17.925049Z",
          "score": 100,
          "model": "gpt-6-1-sol",
          "runner": "benchmark_agent",
          "config": "openai-gpt-6-1-sol-max-provider-adapter",
          "verified_publish": true,
          "levels_completed": 8,
          "actions": 253,
          "resets": 0,
          "state": "WIN",
          "completed": true,
          "number_of_levels": 0,
          "level_scores": [
            115,
            115,
            115,
            115,
            115,
            115,
            115,
            115
          ],
          "level_actions": [
            13,
            26,
            41,
            30,
            29,
            25,
            50,
            39
          ],
          "level_baseline_actions": [
            20,
            89,
            106,
            54,
            162,
            38,
            86,
            83
          ]
        },
        {
          "session_id": "08c987ca-9615-4491-9f66-c9c93e411dfa",
          "game_id": "s5i5",
          "game_version": "18d95033",
          "published_at": "2026-09-15T22:30:07.609425Z",
          "score": 100,
          "model": "gpt-6-1-sol",
          "runner": "benchmark_agent",
          "config": "openai-gpt-6-1-sol-max",
          "verified_publish": true,
          "levels_completed": 8,
          "actions": 261,
          "resets": 0,
          "state": "WIN",
          "completed": true,
          "number_of_levels": 0,
          "level_scores": [
            115,
            115,
            115,
            115,
            115,
            115,
            115,
            115
          ],
          "level_actions": [
            14,
            26,
            40,
            30,
            30,
            27,
            50,
            44
          ],
          "level_baseline_actions": [
            20,
            89,
            106,
            54,
            162,
            38,
            86,
            83
          ]
        },
        {
          "session_id": "df2679fd-38fc-49be-a1bb-ad9c24f6ef52",
          "game_id": "s5i5",
          "game_version": "18d95033",
          "published_at": "2026-09-12T13:50:23.094739Z",
          "score": 77.77777777777779,
          "model": "gemini-3.8-flash",
          "runner": "benchmark_agent",
          "config": "google-gemini-3-8-flash-high",
          "verified_publish": true,
          "levels_completed": 7,
          "actions": 643,
          "resets": 7,
          "state": "NOT_FINISHED",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            115,
            115,
            115,
            115,
            115,
            115,
            62.25065230199479,
            0
          ],
          "level_actions": [
            13,
            32,
            95,
            30,
            115,
            32,
            109,
            217
          ],
          "level_baseline_actions": [
            20,
            89,
            106,
            54,
            162,
            38,
            86,
            83
          ]
        },
        {
          "session_id": "519894f3-e4e5-4dda-b118-8df1ab52b814",
          "game_id": "s5i5",
          "game_version": "18d95033",
          "published_at": "2026-09-12T11:52:54.27266Z",
          "score": 7.76106065469373,
          "model": "gemini-3.8-flash",
          "runner": "benchmark_agent",
          "config": "google-gemini-3-8-flash-low",
          "verified_publish": true,
          "levels_completed": 3,
          "actions": 731,
          "resets": 6,
          "state": "GAME_OVER",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            26.29848783694937,
            115,
            7.699898577341631,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            39,
            40,
            382,
            270,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            20,
            89,
            106,
            54,
            162,
            38,
            86,
            83
          ]
        },
        {
          "session_id": "47e052d4-6521-4a7f-9d78-6bb8b2d00d18",
          "game_id": "s5i5",
          "game_version": "18d95033",
          "published_at": "2026-09-09T17:34:07.863669Z",
          "score": 8.333333333333332,
          "model": "gemini-3.8-flash",
          "runner": "benchmark_agent",
          "config": "google-gemini-3-8-flash-medium",
          "verified_publish": true,
          "levels_completed": 2,
          "actions": 163,
          "resets": 0,
          "state": "NOT_FINISHED",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            115,
            115,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            17,
            50,
            96,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            20,
            89,
            106,
            54,
            162,
            38,
            86,
            83
          ]
        },
        {
          "session_id": "cfd3c696-e083-4556-a132-4276ebf46763",
          "game_id": "s5i5",
          "game_version": "18d95033",
          "published_at": "2026-08-11T06:09:40.106225Z",
          "score": 0,
          "model": "grok-4.6",
          "runner": "benchmark_agent",
          "config": "xai-grok-4-6-xhigh",
          "verified_publish": true,
          "levels_completed": 0,
          "actions": 100,
          "resets": 1,
          "state": "GAME_OVER",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            0,
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            100,
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            20,
            89,
            106,
            54,
            162,
            38,
            86,
            83
          ]
        },
        {
          "session_id": "cb8fbc91-0b92-441e-a13f-eaccbc0ac86e",
          "game_id": "s5i5",
          "game_version": "18d95033",
          "published_at": "2026-07-21T07:58:07.181643Z",
          "score": 100,
          "model": "claude-opus-5",
          "runner": "benchmark_agent",
          "config": "anthropic-claude-opus-5-high",
          "verified_publish": true,
          "levels_completed": 8,
          "actions": 388,
          "resets": 1,
          "state": "WIN",
          "completed": true,
          "number_of_levels": 0,
          "level_scores": [
            44.44444444444444,
            115,
            115,
            71.19140625,
            115,
            115,
            115,
            91.01598625974368
          ],
          "level_actions": [
            30,
            26,
            39,
            64,
            36,
            33,
            73,
            87
          ],
          "level_baseline_actions": [
            20,
            89,
            106,
            54,
            162,
            38,
            86,
            83
          ]
        },
        {
          "session_id": "c8c1c1e2-17e0-4bc6-9a31-35d7b811add4",
          "game_id": "s5i5",
          "game_version": "18d95033",
          "published_at": "2026-06-24T10:47:49.333293Z",
          "score": 0,
          "model": "gpt-5-6-luna",
          "runner": "benchmark_agent",
          "config": "openai-gpt-5-6-luna-low",
          "verified_publish": true,
          "levels_completed": 0,
          "actions": 100,
          "resets": 2,
          "state": "GAME_OVER",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            0,
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            100,
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            20,
            89,
            106,
            54,
            162,
            38,
            86,
            83
          ]
        },
        {
          "session_id": "370eaf10-88bb-4300-8d05-1da4f369d6dd",
          "game_id": "s5i5",
          "game_version": "18d95033",
          "published_at": "2026-06-24T10:27:48.26311Z",
          "score": 0,
          "model": "gpt-5-6-luna",
          "runner": "benchmark_agent",
          "config": "openai-gpt-5-6-luna-medium",
          "verified_publish": true,
          "levels_completed": 0,
          "actions": 100,
          "resets": 3,
          "state": "GAME_OVER",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            0,
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            100,
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            20,
            89,
            106,
            54,
            162,
            38,
            86,
            83
          ]
        },
        {
          "session_id": "067cf187-ba61-49c9-b53b-7162af823604",
          "game_id": "s5i5",
          "game_version": "18d95033",
          "published_at": "2026-06-24T03:27:26.283596Z",
          "score": 0,
          "model": "gpt-5-6-luna",
          "runner": "benchmark_agent",
          "config": "openai-gpt-5-6-luna-high",
          "verified_publish": true,
          "levels_completed": 0,
          "actions": 100,
          "resets": 7,
          "state": "NOT_FINISHED",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            0,
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            100,
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            20,
            89,
            106,
            54,
            162,
            38,
            86,
            83
          ]
        },
        {
          "session_id": "3c27c3af-16c8-4063-8136-d333e08ba8a4",
          "game_id": "s5i5",
          "game_version": "18d95033",
          "published_at": "2026-06-24T01:22:48.803003Z",
          "score": 0,
          "model": "gpt-5-6-luna",
          "runner": "benchmark_agent",
          "config": "openai-gpt-5-6-luna-xhigh",
          "verified_publish": true,
          "levels_completed": 0,
          "actions": 100,
          "resets": 4,
          "state": "GAME_OVER",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            0,
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            100,
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            20,
            89,
            106,
            54,
            162,
            38,
            86,
            83
          ]
        },
        {
          "session_id": "06b27277-21e3-45cd-a689-86ee372ae3b4",
          "game_id": "s5i5",
          "game_version": "18d95033",
          "published_at": "2026-06-23T03:55:21.089424Z",
          "score": 0,
          "model": "gpt-5-6-luna",
          "runner": "benchmark_agent",
          "config": "openai-gpt-5-6-luna-max",
          "verified_publish": true,
          "levels_completed": 0,
          "actions": 100,
          "resets": 17,
          "state": "NOT_FINISHED",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            0,
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            100,
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            20,
            89,
            106,
            54,
            162,
            38,
            86,
            83
          ]
        },
        {
          "session_id": "4e16e7a6-bc3b-4e6f-b5fb-14b70e303be6",
          "game_id": "s5i5",
          "game_version": "18d95033",
          "published_at": "2026-06-20T11:55:08.162757Z",
          "score": 0,
          "model": "gpt-5-6-terra",
          "runner": "benchmark_agent",
          "config": "openai-gpt-5-6-terra-low",
          "verified_publish": true,
          "levels_completed": 0,
          "actions": 80,
          "resets": 2,
          "state": "GAME_OVER",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            0,
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            80,
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            20,
            89,
            106,
            54,
            162,
            38,
            86,
            83
          ]
        },
        {
          "session_id": "605e647f-35d5-440e-9230-dc6fbe59fa0c",
          "game_id": "s5i5",
          "game_version": "18d95033",
          "published_at": "2026-06-20T09:12:21.585524Z",
          "score": 0,
          "model": "gpt-5-6-terra",
          "runner": "benchmark_agent",
          "config": "openai-gpt-5-6-terra-medium",
          "verified_publish": true,
          "levels_completed": 0,
          "actions": 100,
          "resets": 7,
          "state": "NOT_FINISHED",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            0,
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            100,
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            20,
            89,
            106,
            54,
            162,
            38,
            86,
            83
          ]
        },
        {
          "session_id": "0293607f-c47a-402d-8c60-e178c51c07c7",
          "game_id": "s5i5",
          "game_version": "18d95033",
          "published_at": "2026-06-20T08:34:59.046111Z",
          "score": 1.1562030292519365,
          "model": "gpt-5-6-terra",
          "runner": "benchmark_agent",
          "config": "openai-gpt-5-6-terra-xhigh",
          "verified_publish": true,
          "levels_completed": 1,
          "actions": 476,
          "resets": 18,
          "state": "NOT_FINISHED",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            41.62330905306972,
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            31,
            445,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            20,
            89,
            106,
            54,
            162,
            38,
            86,
            83
          ]
        },
        {
          "session_id": "e9089b50-abc4-403c-93c3-82643ff5b3ea",
          "game_id": "s5i5",
          "game_version": "18d95033",
          "published_at": "2026-06-20T08:20:26.27445Z",
          "score": 0,
          "model": "gpt-5-6-terra",
          "runner": "benchmark_agent",
          "config": "openai-gpt-5-6-terra-high",
          "verified_publish": true,
          "levels_completed": 0,
          "actions": 100,
          "resets": 3,
          "state": "NOT_FINISHED",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            0,
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            100,
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            20,
            89,
            106,
            54,
            162,
            38,
            86,
            83
          ]
        },
        {
          "session_id": "01d32bd2-a579-4b85-bda8-0dda4a8362e1",
          "game_id": "s5i5",
          "game_version": "18d95033",
          "published_at": "2026-06-19T08:19:54.175514Z",
          "score": 6.5501766100546766,
          "model": "gpt-5-6-terra",
          "runner": "benchmark_agent",
          "config": "openai-gpt-5-6-terra-max",
          "verified_publish": true,
          "levels_completed": 2,
          "actions": 671,
          "resets": 80,
          "state": "NOT_FINISHED",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            5.806357961968355,
            115,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            83,
            58,
            530,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            20,
            89,
            106,
            54,
            162,
            38,
            86,
            83
          ]
        },
        {
          "session_id": "ec63e8a9-22c0-4d3c-8e37-ace217def1be",
          "game_id": "s5i5",
          "game_version": "18d95033",
          "published_at": "2026-06-15T16:44:37.555751Z",
          "score": 0,
          "model": "gpt-5-6-sol",
          "runner": "benchmark_agent",
          "config": "openai-gpt-5-6-sol-xhigh",
          "verified_publish": true,
          "levels_completed": 0,
          "actions": 100,
          "resets": 43,
          "state": "NOT_FINISHED",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            0,
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            100,
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            20,
            89,
            106,
            54,
            162,
            38,
            86,
            83
          ]
        },
        {
          "session_id": "05c40339-f4f8-4f2f-8d75-e47c6d1faa36",
          "game_id": "s5i5",
          "game_version": "18d95033",
          "published_at": "2026-06-13T01:54:18.258605Z",
          "score": 0,
          "model": "gpt-5-6-sol",
          "runner": "benchmark_agent",
          "config": "openai-gpt-5-6-sol-max",
          "verified_publish": true,
          "levels_completed": 0,
          "actions": 100,
          "resets": 29,
          "state": "NOT_FINISHED",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            0,
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            100,
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            20,
            89,
            106,
            54,
            162,
            38,
            86,
            83
          ]
        },
        {
          "session_id": "62e3b696-74e3-4e5f-afba-9ab54dd9d6eb",
          "game_id": "s5i5",
          "game_version": "18d95033",
          "published_at": "2026-06-10T10:53:04.563611Z",
          "score": 0,
          "model": "gpt-5-6-sol",
          "runner": "benchmark_agent",
          "config": "openai-gpt-5-6-sol-high",
          "verified_publish": true,
          "levels_completed": 0,
          "actions": 100,
          "resets": 38,
          "state": "NOT_FINISHED",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            0,
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            100,
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            20,
            89,
            106,
            54,
            162,
            38,
            86,
            83
          ]
        },
        {
          "session_id": "cfdc1ef6-20d8-4b72-a0ba-439d03734ae9",
          "game_id": "s5i5",
          "game_version": "18d95033",
          "published_at": "2026-06-10T10:34:36.550857Z",
          "score": 0.7694675284702984,
          "model": "gpt-5-6-sol",
          "runner": "benchmark_agent",
          "config": "openai-gpt-5-6-sol-medium",
          "verified_publish": true,
          "levels_completed": 1,
          "actions": 483,
          "resets": 3,
          "state": "GAME_OVER",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            27.700831024930743,
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            38,
            445,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            20,
            89,
            106,
            54,
            162,
            38,
            86,
            83
          ]
        },
        {
          "session_id": "0d984880-5c82-462a-85f3-461197327a09",
          "game_id": "s5i5",
          "game_version": "18d95033",
          "published_at": "2026-06-09T21:41:11.10454Z",
          "score": 0,
          "model": "gpt-5-6-sol",
          "runner": "benchmark_agent",
          "config": "openai-gpt-5-6-sol-low",
          "verified_publish": true,
          "levels_completed": 0,
          "actions": 100,
          "resets": 3,
          "state": "GAME_OVER",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            0,
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            100,
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            20,
            89,
            106,
            54,
            162,
            38,
            86,
            83
          ]
        },
        {
          "session_id": "fb4d7bea-71f8-4f65-b31b-6a8027a79b90",
          "game_id": "s5i5",
          "game_version": "18d95033",
          "published_at": "2026-05-29T22:35:48.759673Z",
          "score": 5.932506887052341,
          "model": "claude-opus-4-8",
          "runner": "benchmark_agent",
          "config": "anthropic-opus-4-8-high",
          "verified_publish": true,
          "levels_completed": 2,
          "actions": 309,
          "resets": 5,
          "state": "NOT_FINISHED",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            82.64462809917354,
            65.46280991735537,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            22,
            110,
            177,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            20,
            89,
            106,
            54,
            162,
            38,
            86,
            83
          ]
        },
        {
          "session_id": "cb612ce5-c0ea-4007-b7ae-f2cb9c114c65",
          "game_id": "s5i5",
          "game_version": "18d95033",
          "published_at": "2026-05-01T05:07:55.472664Z",
          "score": 0,
          "model": "claude-opus-4-7",
          "runner": "benchmark_agent",
          "config": "anthropic-opus-4-7-high",
          "verified_publish": true,
          "levels_completed": 0,
          "actions": 100,
          "resets": 9,
          "state": "NOT_FINISHED",
          "completed": false,
          "number_of_levels": 8,
          "level_scores": [
            0,
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            100,
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            20,
            89,
            106,
            54,
            162,
            38,
            86,
            83
          ]
        },
        {
          "session_id": "9ac09abc-1ff4-4d6b-86fc-741a18436bda",
          "game_id": "s5i5",
          "game_version": "18d95033",
          "published_at": "2026-04-30T23:24:40.150163Z",
          "score": 0,
          "model": "gpt-5.5-2026-04-23",
          "runner": "benchmark_agent",
          "config": "openai-gpt-5-5-2026-04-23-high",
          "verified_publish": true,
          "levels_completed": 0,
          "actions": 100,
          "resets": 2,
          "state": "GAME_OVER",
          "completed": false,
          "number_of_levels": 8,
          "level_scores": [
            0,
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            100,
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            20,
            89,
            106,
            54,
            162,
            38,
            86,
            83
          ]
        },
        {
          "session_id": "c3e49182-d874-4c34-9cd6-454cd2679ebb",
          "game_id": "s5i5",
          "game_version": "18d95033",
          "published_at": "2026-04-14T17:43:59.519012Z",
          "score": 0,
          "model": "grok-4.20-beta-0309-reasoning",
          "runner": "benchmark_agent",
          "config": "xai-grok-4-20-beta-0309-reasoning",
          "verified_publish": true,
          "levels_completed": 0,
          "actions": 100,
          "resets": 2,
          "state": "GAME_OVER",
          "completed": false,
          "number_of_levels": 8,
          "level_scores": [
            0,
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            100,
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            20,
            89,
            106,
            54,
            162,
            38,
            86,
            83
          ]
        },
        {
          "session_id": "c8733fc2-2b86-4e3b-98a4-ed9bb46a2342",
          "game_id": "s5i5",
          "game_version": "18d95033",
          "published_at": "2026-04-14T17:43:59.184422Z",
          "score": 0,
          "model": "gpt-5.4-2026-03-05",
          "runner": "benchmark_agent",
          "config": "openai-gpt-5-4-2026-03-05-high",
          "verified_publish": true,
          "levels_completed": 0,
          "actions": 100,
          "resets": 2,
          "state": "GAME_OVER",
          "completed": false,
          "number_of_levels": 8,
          "level_scores": [
            0,
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            100,
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            20,
            89,
            106,
            54,
            162,
            38,
            86,
            83
          ]
        },
        {
          "session_id": "343a61de-2302-4dae-b954-f24187fb75cc",
          "game_id": "s5i5",
          "game_version": "18d95033",
          "published_at": "2026-04-14T17:43:58.929154Z",
          "score": 2.7777777777777777,
          "model": "gemini-3.1-pro-preview",
          "runner": "benchmark_agent",
          "config": "google-gemini-3-1-pro-preview",
          "verified_publish": true,
          "levels_completed": 1,
          "actions": 18,
          "resets": 0,
          "state": "NOT_FINISHED",
          "completed": false,
          "number_of_levels": 8,
          "level_scores": [
            115,
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            18,
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            20,
            89,
            106,
            54,
            162,
            38,
            86,
            83
          ]
        },
        {
          "session_id": "203183b5-eeb3-488d-b226-32963bddc950",
          "game_id": "s5i5",
          "game_version": "18d95033",
          "published_at": "2026-04-14T17:43:58.542891Z",
          "score": 0,
          "model": "claude-opus-4-6",
          "runner": "benchmark_agent",
          "config": "anthropic-opus-4-6-max-effort",
          "verified_publish": true,
          "levels_completed": 0,
          "actions": 51,
          "resets": 1,
          "state": "GAME_OVER",
          "completed": false,
          "number_of_levels": 8,
          "level_scores": [
            0,
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            51,
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            20,
            89,
            106,
            54,
            162,
            38,
            86,
            83
          ]
        }
      ]
    },
    {
      "game": "ka59",
      "source": "https://arcprize.org/tasks/ka59",
      "runs": [
        {
          "session_id": "839e4f0a-d10e-4fbc-b59d-3183b6abc8cf",
          "game_id": "ka59",
          "game_version": "38d34dbb",
          "published_at": "2026-09-25T15:27:54.473855Z",
          "score": 1.7499999999999998,
          "model": "gpt-6-sol",
          "runner": "benchmark_agent",
          "config": "openai-gpt-6-sol-max",
          "verified_publish": true,
          "levels_completed": 1,
          "actions": 346,
          "resets": 3,
          "state": "GAME_OVER",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            48.99999999999999,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            40,
            306,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            28,
            109,
            51,
            51,
            33,
            132,
            326
          ]
        },
        {
          "session_id": "6fc3a5cb-817b-463c-a8c2-2eef4b3104e5",
          "game_id": "ka59",
          "game_version": "38d34dbb",
          "published_at": "2026-09-25T11:50:55.989659Z",
          "score": 3.571428571428571,
          "model": "gpt-6-sol",
          "runner": "benchmark_agent",
          "config": "openai-gpt-6-sol-xhigh",
          "verified_publish": true,
          "levels_completed": 1,
          "actions": 569,
          "resets": 4,
          "state": "GAME_OVER",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            115,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            24,
            545,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            28,
            109,
            51,
            51,
            33,
            132,
            326
          ]
        },
        {
          "session_id": "20a046a3-a99c-4247-8616-7d99c366d1c5",
          "game_id": "ka59",
          "game_version": "38d34dbb",
          "published_at": "2026-09-25T08:50:06.228549Z",
          "score": 0.7054673721340388,
          "model": "gpt-6-sol",
          "runner": "benchmark_agent",
          "config": "openai-gpt-6-sol-high",
          "verified_publish": true,
          "levels_completed": 1,
          "actions": 608,
          "resets": 4,
          "state": "GAME_OVER",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            19.753086419753085,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            63,
            545,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            28,
            109,
            51,
            51,
            33,
            132,
            326
          ]
        },
        {
          "session_id": "5e0dc34c-e3ff-4997-8e4e-dca1c9ab3ba4",
          "game_id": "ka59",
          "game_version": "38d34dbb",
          "published_at": "2026-09-25T07:15:58.019208Z",
          "score": 0.18210197710717999,
          "model": "gpt-6-sol",
          "runner": "benchmark_agent",
          "config": "openai-gpt-6-sol-none-provider-adapter",
          "verified_publish": true,
          "levels_completed": 1,
          "actions": 669,
          "resets": 7,
          "state": "GAME_OVER",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            5.09885535900104,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            124,
            545,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            28,
            109,
            51,
            51,
            33,
            132,
            326
          ]
        },
        {
          "session_id": "832e0b8a-1d63-4e36-826d-71e4bbf09a16",
          "game_id": "ka59",
          "game_version": "38d34dbb",
          "published_at": "2026-09-25T07:01:24.665782Z",
          "score": 2.422145328719723,
          "model": "gpt-6-sol",
          "runner": "benchmark_agent",
          "config": "openai-gpt-6-sol-medium",
          "verified_publish": true,
          "levels_completed": 1,
          "actions": 579,
          "resets": 4,
          "state": "GAME_OVER",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            67.82006920415225,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            34,
            545,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            28,
            109,
            51,
            51,
            33,
            132,
            326
          ]
        },
        {
          "session_id": "f102bc3b-6e4a-4168-8e37-46ac34f25ddc",
          "game_id": "ka59",
          "game_version": "38d34dbb",
          "published_at": "2026-09-25T06:59:56.954806Z",
          "score": 4.704236242126116,
          "model": "gpt-6-sol",
          "runner": "benchmark_agent",
          "config": "openai-gpt-6-sol-low-provider-adapter",
          "verified_publish": true,
          "levels_completed": 2,
          "actions": 657,
          "resets": 5,
          "state": "GAME_OVER",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            115,
            8.359307389765636,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            25,
            377,
            255,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            28,
            109,
            51,
            51,
            33,
            132,
            326
          ]
        },
        {
          "session_id": "32e00da1-7b1c-4999-803d-a47e395adf5a",
          "game_id": "ka59",
          "game_version": "38d34dbb",
          "published_at": "2026-09-25T06:47:31.036307Z",
          "score": 2.734375,
          "model": "gpt-6-sol",
          "runner": "benchmark_agent",
          "config": "openai-gpt-6-sol-none",
          "verified_publish": true,
          "levels_completed": 1,
          "actions": 577,
          "resets": 4,
          "state": "GAME_OVER",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            76.5625,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            32,
            545,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            28,
            109,
            51,
            51,
            33,
            132,
            326
          ]
        },
        {
          "session_id": "e00a41aa-00a6-48e0-a90c-31af221947c4",
          "game_id": "ka59",
          "game_version": "38d34dbb",
          "published_at": "2026-09-25T06:40:35.168911Z",
          "score": 3.571428571428571,
          "model": "gpt-6-sol",
          "runner": "benchmark_agent",
          "config": "openai-gpt-6-sol-medium-provider-adapter",
          "verified_publish": true,
          "levels_completed": 1,
          "actions": 570,
          "resets": 6,
          "state": "GAME_OVER",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            115,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            25,
            545,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            28,
            109,
            51,
            51,
            33,
            132,
            326
          ]
        },
        {
          "session_id": "b45ad071-dff9-4bf4-bda6-db8f39804bb3",
          "game_id": "ka59",
          "game_version": "38d34dbb",
          "published_at": "2026-09-25T06:09:49.535616Z",
          "score": 8.453940181737183,
          "model": "gpt-6-sol",
          "runner": "benchmark_agent",
          "config": "openai-gpt-6-sol-high-provider-adapter",
          "verified_publish": true,
          "levels_completed": 3,
          "actions": 672,
          "resets": 7,
          "state": "NOT_FINISHED",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            115,
            12.688495877653894,
            32.11111111111111,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            21,
            306,
            90,
            255,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            28,
            109,
            51,
            51,
            33,
            132,
            326
          ]
        },
        {
          "session_id": "fc9b0160-919d-4100-9273-af025b39f07e",
          "game_id": "ka59",
          "game_version": "38d34dbb",
          "published_at": "2026-09-25T06:00:17.271Z",
          "score": 0,
          "model": "gpt-6-sol",
          "runner": "benchmark_agent",
          "config": "openai-gpt-6-sol-low",
          "verified_publish": true,
          "levels_completed": 0,
          "actions": 140,
          "resets": 1,
          "state": "NOT_FINISHED",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            140,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            28,
            109,
            51,
            51,
            33,
            132,
            326
          ]
        },
        {
          "session_id": "2965ecf0-208f-4327-bdf7-5c40f579ce6b",
          "game_id": "ka59",
          "game_version": "38d34dbb",
          "published_at": "2026-09-25T05:43:40.611071Z",
          "score": 3.571428571428571,
          "model": "gpt-6-sol",
          "runner": "benchmark_agent",
          "config": "openai-gpt-6-sol-xhigh-provider-adapter",
          "verified_publish": true,
          "levels_completed": 1,
          "actions": 565,
          "resets": 8,
          "state": "NOT_FINISHED",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            115,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            20,
            545,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            28,
            109,
            51,
            51,
            33,
            132,
            326
          ]
        },
        {
          "session_id": "a6fdd583-2572-4a86-9317-2439a2ff9a58",
          "game_id": "ka59",
          "game_version": "38d34dbb",
          "published_at": "2026-09-21T10:21:11.884412Z",
          "score": 12.53120389509697,
          "model": "gemini-3.8-flash",
          "runner": "benchmark_agent",
          "config": "google-gemini-3-8-flash-low-provider-adapter",
          "verified_publish": true,
          "levels_completed": 4,
          "actions": 597,
          "resets": 6,
          "state": "GAME_OVER",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            24.13050169282856,
            115,
            21.892096624863232,
            7.76672937382424,
            0,
            0,
            0
          ],
          "level_actions": [
            57,
            83,
            109,
            183,
            165,
            0,
            0
          ],
          "level_baseline_actions": [
            28,
            109,
            51,
            51,
            33,
            132,
            326
          ]
        },
        {
          "session_id": "809e0821-82f5-4511-af75-1328e2bc409c",
          "game_id": "ka59",
          "game_version": "38d34dbb",
          "published_at": "2026-09-20T11:02:38.746654Z",
          "score": 25.76459193112737,
          "model": "gemini-3.8-flash",
          "runner": "benchmark_agent",
          "config": "google-gemini-3-8-flash-medium-provider-adapter",
          "verified_publish": true,
          "levels_completed": 6,
          "actions": 2332,
          "resets": 13,
          "state": "GAME_OVER",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            115,
            58.10064061812312,
            63.5009765625,
            12.543402777777779,
            15.072664359861593,
            29.027905039566846,
            0
          ],
          "level_actions": [
            21,
            143,
            64,
            144,
            85,
            245,
            1630
          ],
          "level_baseline_actions": [
            28,
            109,
            51,
            51,
            33,
            132,
            326
          ]
        },
        {
          "session_id": "bdf11c19-e895-4114-8805-5a02ec747321",
          "game_id": "ka59",
          "game_version": "38d34dbb",
          "published_at": "2026-09-17T19:14:37.943344Z",
          "score": 14.853345718008743,
          "model": "gemini-3.8-flash",
          "runner": "benchmark_agent",
          "config": "google-gemini-3-8-flash-high-provider-adapter",
          "verified_publish": true,
          "levels_completed": 3,
          "actions": 811,
          "resets": 8,
          "state": "GAME_OVER",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            60.49382716049383,
            5.199926471875493,
            115,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            36,
            478,
            42,
            255,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            28,
            109,
            51,
            51,
            33,
            132,
            326
          ]
        },
        {
          "session_id": "5bf6abde-dd86-44df-9cb2-d1ae15f864ef",
          "game_id": "ka59",
          "game_version": "38d34dbb",
          "published_at": "2026-09-16T19:31:22.729161Z",
          "score": 0,
          "model": "gpt-6-luna",
          "runner": "benchmark_agent",
          "config": "openai-gpt-6-luna-medium-provider-adapter",
          "verified_publish": true,
          "levels_completed": 0,
          "actions": 140,
          "resets": 3,
          "state": "NOT_FINISHED",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            140,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            28,
            109,
            51,
            51,
            33,
            132,
            326
          ]
        },
        {
          "session_id": "affe4ef9-f784-4784-9a6a-bd5ac32363fe",
          "game_id": "ka59",
          "game_version": "38d34dbb",
          "published_at": "2026-09-16T19:26:12.255209Z",
          "score": 0,
          "model": "gpt-6-luna",
          "runner": "benchmark_agent",
          "config": "openai-gpt-6-luna-low",
          "verified_publish": true,
          "levels_completed": 0,
          "actions": 140,
          "resets": 6,
          "state": "GAME_OVER",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            140,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            28,
            109,
            51,
            51,
            33,
            132,
            326
          ]
        },
        {
          "session_id": "d6daacda-cf0e-4615-987f-676f2c01d90e",
          "game_id": "ka59",
          "game_version": "38d34dbb",
          "published_at": "2026-09-16T19:25:38.7267Z",
          "score": 0,
          "model": "gpt-6-luna",
          "runner": "benchmark_agent",
          "config": "openai-gpt-6-luna-low-provider-adapter",
          "verified_publish": true,
          "levels_completed": 0,
          "actions": 140,
          "resets": 6,
          "state": "NOT_FINISHED",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            140,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            28,
            109,
            51,
            51,
            33,
            132,
            326
          ]
        },
        {
          "session_id": "c4cb8557-40f3-467f-a5b5-d5b2391251a9",
          "game_id": "ka59",
          "game_version": "38d34dbb",
          "published_at": "2026-09-16T19:24:56.81807Z",
          "score": 0,
          "model": "gpt-6-luna",
          "runner": "benchmark_agent",
          "config": "openai-gpt-6-luna-none-provider-adapter",
          "verified_publish": true,
          "levels_completed": 0,
          "actions": 140,
          "resets": 4,
          "state": "NOT_FINISHED",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            140,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            28,
            109,
            51,
            51,
            33,
            132,
            326
          ]
        },
        {
          "session_id": "9a390831-15be-4ed2-b74a-7d707bd8b39a",
          "game_id": "ka59",
          "game_version": "38d34dbb",
          "published_at": "2026-09-16T19:24:05.756811Z",
          "score": 0,
          "model": "gpt-6-luna",
          "runner": "benchmark_agent",
          "config": "openai-gpt-6-luna-none",
          "verified_publish": true,
          "levels_completed": 0,
          "actions": 140,
          "resets": 8,
          "state": "NOT_FINISHED",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            140,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            28,
            109,
            51,
            51,
            33,
            132,
            326
          ]
        },
        {
          "session_id": "f330a9b6-06aa-4d2d-88ca-3df41c8bd739",
          "game_id": "ka59",
          "game_version": "38d34dbb",
          "published_at": "2026-09-16T17:32:09.326536Z",
          "score": 40.036153353087336,
          "model": "gpt-6-1-sol",
          "runner": "benchmark_agent",
          "config": "openai-gpt-6-1-sol-high",
          "verified_publish": true,
          "levels_completed": 5,
          "actions": 1004,
          "resets": 12,
          "state": "NOT_FINISHED",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            87.11111111111111,
            42.09537981859411,
            104.03999999999999,
            69.90056436441817,
            71.59763313609467,
            0,
            0
          ],
          "level_actions": [
            30,
            168,
            50,
            61,
            39,
            656,
            0
          ],
          "level_baseline_actions": [
            28,
            109,
            51,
            51,
            33,
            132,
            326
          ]
        },
        {
          "session_id": "be0f8023-d16d-4e68-9062-3f57bd8f1c6f",
          "game_id": "ka59",
          "game_version": "38d34dbb",
          "published_at": "2026-09-16T15:00:13.831127Z",
          "score": 75,
          "model": "gpt-6-1-sol",
          "runner": "benchmark_agent",
          "config": "openai-gpt-6-1-sol-max",
          "verified_publish": true,
          "levels_completed": 6,
          "actions": 870,
          "resets": 8,
          "state": "NOT_FINISHED",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            115,
            115,
            115,
            46.24000000000001,
            115,
            115,
            0
          ],
          "level_actions": [
            22,
            50,
            35,
            75,
            27,
            94,
            567
          ],
          "level_baseline_actions": [
            28,
            109,
            51,
            51,
            33,
            132,
            326
          ]
        },
        {
          "session_id": "6a10a9af-906d-4655-b4ed-437b119e5423",
          "game_id": "ka59",
          "game_version": "38d34dbb",
          "published_at": "2026-09-16T08:56:47.441407Z",
          "score": 0,
          "model": "gpt-6-luna",
          "runner": "benchmark_agent",
          "config": "openai-gpt-6-luna-medium",
          "verified_publish": true,
          "levels_completed": 0,
          "actions": 140,
          "resets": 6,
          "state": "NOT_FINISHED",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            140,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            28,
            109,
            51,
            51,
            33,
            132,
            326
          ]
        },
        {
          "session_id": "374c5c75-daea-44c5-88f7-fc0555472bd7",
          "game_id": "ka59",
          "game_version": "38d34dbb",
          "published_at": "2026-09-16T08:40:39.988834Z",
          "score": 0,
          "model": "gpt-6-luna",
          "runner": "benchmark_agent",
          "config": "openai-gpt-6-luna-high-provider-adapter",
          "verified_publish": true,
          "levels_completed": 0,
          "actions": 140,
          "resets": 1,
          "state": "GAME_OVER",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            140,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            28,
            109,
            51,
            51,
            33,
            132,
            326
          ]
        },
        {
          "session_id": "f36ab1ef-0229-4ffe-815d-b8135a331def",
          "game_id": "ka59",
          "game_version": "38d34dbb",
          "published_at": "2026-09-16T08:37:46.706932Z",
          "score": 0,
          "model": "gpt-6-luna",
          "runner": "benchmark_agent",
          "config": "openai-gpt-6-luna-high",
          "verified_publish": true,
          "levels_completed": 0,
          "actions": 140,
          "resets": 3,
          "state": "NOT_FINISHED",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            140,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            28,
            109,
            51,
            51,
            33,
            132,
            326
          ]
        },
        {
          "session_id": "b0b571fb-e208-4eef-a51f-a2a3b3fae34c",
          "game_id": "ka59",
          "game_version": "38d34dbb",
          "published_at": "2026-09-16T07:51:51.869651Z",
          "score": 0,
          "model": "gpt-6-luna",
          "runner": "benchmark_agent",
          "config": "openai-gpt-6-luna-xhigh-provider-adapter",
          "verified_publish": true,
          "levels_completed": 0,
          "actions": 140,
          "resets": 1,
          "state": "NOT_FINISHED",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            140,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            28,
            109,
            51,
            51,
            33,
            132,
            326
          ]
        },
        {
          "session_id": "ba8f1b73-dad6-4cb8-809a-955bf624fa70",
          "game_id": "ka59",
          "game_version": "38d34dbb",
          "published_at": "2026-09-16T07:50:35.587497Z",
          "score": 45.11158708976908,
          "model": "gpt-6-1-sol",
          "runner": "benchmark_agent",
          "config": "openai-gpt-6-1-sol-xhigh",
          "verified_publish": true,
          "levels_completed": 5,
          "actions": 1034,
          "resets": 10,
          "state": "NOT_FINISHED",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            107.54458161865568,
            27.72760157763308,
            115,
            45.03116343490304,
            115,
            0,
            0
          ],
          "level_actions": [
            27,
            207,
            34,
            76,
            30,
            660,
            0
          ],
          "level_baseline_actions": [
            28,
            109,
            51,
            51,
            33,
            132,
            326
          ]
        },
        {
          "session_id": "c4117b57-a164-4b01-872a-0848696c43d7",
          "game_id": "ka59",
          "game_version": "38d34dbb",
          "published_at": "2026-09-16T07:33:25.308704Z",
          "score": 0,
          "model": "gpt-6-luna",
          "runner": "benchmark_agent",
          "config": "openai-gpt-6-luna-xhigh",
          "verified_publish": true,
          "levels_completed": 0,
          "actions": 140,
          "resets": 3,
          "state": "NOT_FINISHED",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            140,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            28,
            109,
            51,
            51,
            33,
            132,
            326
          ]
        },
        {
          "session_id": "35e1c2ce-0ea9-40e6-b461-8a0ed5aa12aa",
          "game_id": "ka59",
          "game_version": "38d34dbb",
          "published_at": "2026-09-16T07:10:10.092037Z",
          "score": 0.68359375,
          "model": "gpt-6-luna",
          "runner": "benchmark_agent",
          "config": "openai-gpt-6-luna-max-provider-adapter",
          "verified_publish": true,
          "levels_completed": 1,
          "actions": 609,
          "resets": 4,
          "state": "GAME_OVER",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            19.140625,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            64,
            545,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            28,
            109,
            51,
            51,
            33,
            132,
            326
          ]
        },
        {
          "session_id": "e7812140-e231-4f4a-b5fc-97d58005e664",
          "game_id": "ka59",
          "game_version": "38d34dbb",
          "published_at": "2026-09-16T07:08:41.295123Z",
          "score": 18.91737109044801,
          "model": "gpt-6-1-sol",
          "runner": "benchmark_agent",
          "config": "openai-gpt-6-1-sol-low",
          "verified_publish": true,
          "levels_completed": 3,
          "actions": 423,
          "resets": 2,
          "state": "GAME_OVER",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            115,
            115,
            61.562130177514796,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            17,
            86,
            65,
            255,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            28,
            109,
            51,
            51,
            33,
            132,
            326
          ]
        },
        {
          "session_id": "f0f0683d-3ae8-4440-8339-a0f0d5b325bd",
          "game_id": "ka59",
          "game_version": "38d34dbb",
          "published_at": "2026-09-16T06:01:33.155549Z",
          "score": 9.17914085535284,
          "model": "gpt-6-1-sol",
          "runner": "benchmark_agent",
          "config": "openai-gpt-6-1-sol-medium",
          "verified_publish": true,
          "levels_completed": 3,
          "actions": 720,
          "resets": 7,
          "state": "NOT_FINISHED",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            115,
            8.493951785867484,
            41.67601345938151,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            12,
            374,
            79,
            255,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            28,
            109,
            51,
            51,
            33,
            132,
            326
          ]
        },
        {
          "session_id": "acda9f3c-155f-463b-bdec-7fc4fe681772",
          "game_id": "ka59",
          "game_version": "38d34dbb",
          "published_at": "2026-09-16T04:56:37.420659Z",
          "score": 100,
          "model": "gpt-6-1-sol",
          "runner": "benchmark_agent",
          "config": "openai-gpt-6-1-sol-high-provider-adapter",
          "verified_publish": true,
          "levels_completed": 7,
          "actions": 384,
          "resets": 0,
          "state": "WIN",
          "completed": true,
          "number_of_levels": 0,
          "level_scores": [
            54.293628808864256,
            115,
            115,
            115,
            115,
            115,
            115
          ],
          "level_actions": [
            38,
            47,
            37,
            42,
            27,
            79,
            114
          ],
          "level_baseline_actions": [
            28,
            109,
            51,
            51,
            33,
            132,
            326
          ]
        },
        {
          "session_id": "9c54fa34-5c0d-4009-a88f-6be3d2872f97",
          "game_id": "ka59",
          "game_version": "38d34dbb",
          "published_at": "2026-09-16T04:42:18.909634Z",
          "score": 0,
          "model": "gpt-6-luna",
          "runner": "benchmark_agent",
          "config": "openai-gpt-6-luna-max",
          "verified_publish": true,
          "levels_completed": 0,
          "actions": 140,
          "resets": 9,
          "state": "NOT_FINISHED",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            140,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            28,
            109,
            51,
            51,
            33,
            132,
            326
          ]
        },
        {
          "session_id": "4b465ff8-0d58-4f21-8ec6-a17339c1770b",
          "game_id": "ka59",
          "game_version": "38d34dbb",
          "published_at": "2026-09-16T04:16:44.196801Z",
          "score": 100,
          "model": "gpt-6-1-sol",
          "runner": "benchmark_agent",
          "config": "openai-gpt-6-1-sol-xhigh-provider-adapter",
          "verified_publish": true,
          "levels_completed": 7,
          "actions": 417,
          "resets": 0,
          "state": "WIN",
          "completed": true,
          "number_of_levels": 0,
          "level_scores": [
            115,
            115,
            115,
            115,
            115,
            95.60493827160494,
            115
          ],
          "level_actions": [
            12,
            53,
            41,
            46,
            21,
            135,
            109
          ],
          "level_baseline_actions": [
            28,
            109,
            51,
            51,
            33,
            132,
            326
          ]
        },
        {
          "session_id": "062d1277-7afe-4672-8ddb-30fd87bb0e52",
          "game_id": "ka59",
          "game_version": "38d34dbb",
          "published_at": "2026-09-16T02:05:24.253358Z",
          "score": 100,
          "model": "gpt-6-1-sol",
          "runner": "benchmark_agent",
          "config": "openai-gpt-6-1-sol-max-provider-adapter",
          "verified_publish": true,
          "levels_completed": 7,
          "actions": 332,
          "resets": 1,
          "state": "WIN",
          "completed": true,
          "number_of_levels": 0,
          "level_scores": [
            115,
            115,
            115,
            115,
            115,
            115,
            115
          ],
          "level_actions": [
            19,
            51,
            34,
            42,
            20,
            70,
            96
          ],
          "level_baseline_actions": [
            28,
            109,
            51,
            51,
            33,
            132,
            326
          ]
        },
        {
          "session_id": "fd354a1f-9608-418f-aedb-4e005a85765d",
          "game_id": "ka59",
          "game_version": "38d34dbb",
          "published_at": "2026-09-12T06:59:19.767262Z",
          "score": 3.571428571428571,
          "model": "gemini-3.8-flash",
          "runner": "benchmark_agent",
          "config": "google-gemini-3-8-flash-high",
          "verified_publish": true,
          "levels_completed": 1,
          "actions": 571,
          "resets": 4,
          "state": "GAME_OVER",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            115,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            26,
            545,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            28,
            109,
            51,
            51,
            33,
            132,
            326
          ]
        },
        {
          "session_id": "d06a9be9-72f5-43f6-a001-ab4e05a60d7b",
          "game_id": "ka59",
          "game_version": "38d34dbb",
          "published_at": "2026-09-11T05:51:24.721193Z",
          "score": 3.571428571428571,
          "model": "gemini-3.8-flash",
          "runner": "benchmark_agent",
          "config": "google-gemini-3-8-flash-low",
          "verified_publish": true,
          "levels_completed": 1,
          "actions": 126,
          "resets": 1,
          "state": "NOT_FINISHED",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            115,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            22,
            104,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            28,
            109,
            51,
            51,
            33,
            132,
            326
          ]
        },
        {
          "session_id": "bb97f3e4-e19a-4847-ac90-88861603ce18",
          "game_id": "ka59",
          "game_version": "38d34dbb",
          "published_at": "2026-09-09T16:38:36.043095Z",
          "score": 3.571428571428571,
          "model": "gemini-3.8-flash",
          "runner": "benchmark_agent",
          "config": "google-gemini-3-8-flash-medium",
          "verified_publish": true,
          "levels_completed": 1,
          "actions": 40,
          "resets": 0,
          "state": "NOT_FINISHED",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            115,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            21,
            19,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            28,
            109,
            51,
            51,
            33,
            132,
            326
          ]
        },
        {
          "session_id": "5f694d5d-45e8-406d-9b21-d1ec8ae6f548",
          "game_id": "ka59",
          "game_version": "38d34dbb",
          "published_at": "2026-08-11T15:04:45.245284Z",
          "score": 3.571428571428571,
          "model": "grok-4.6",
          "runner": "benchmark_agent",
          "config": "xai-grok-4-6-xhigh",
          "verified_publish": true,
          "levels_completed": 1,
          "actions": 565,
          "resets": 4,
          "state": "NOT_FINISHED",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            115,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            20,
            545,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            28,
            109,
            51,
            51,
            33,
            132,
            326
          ]
        },
        {
          "session_id": "d1343ad7-11d6-4860-b02f-c55aea648a5f",
          "game_id": "ka59",
          "game_version": "38d34dbb",
          "published_at": "2026-07-21T07:28:45.849506Z",
          "score": 10.425130762980407,
          "model": "claude-opus-5",
          "runner": "benchmark_agent",
          "config": "anthropic-claude-opus-5-high",
          "verified_publish": true,
          "levels_completed": 3,
          "actions": 601,
          "resets": 8,
          "state": "NOT_FINISHED",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            48.99999999999999,
            19.955993012631012,
            67.66389177939647,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            40,
            244,
            62,
            255,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            28,
            109,
            51,
            51,
            33,
            132,
            326
          ]
        },
        {
          "session_id": "5cd726b0-5f89-4d32-8031-72c3d6eea753",
          "game_id": "ka59",
          "game_version": "38d34dbb",
          "published_at": "2026-06-24T10:49:16.184263Z",
          "score": 0,
          "model": "gpt-5-6-luna",
          "runner": "benchmark_agent",
          "config": "openai-gpt-5-6-luna-low",
          "verified_publish": true,
          "levels_completed": 0,
          "actions": 140,
          "resets": 8,
          "state": "NOT_FINISHED",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            140,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            28,
            109,
            51,
            51,
            33,
            132,
            326
          ]
        },
        {
          "session_id": "0e4b25ac-9f0d-4925-acc0-faefde538a41",
          "game_id": "ka59",
          "game_version": "38d34dbb",
          "published_at": "2026-06-24T10:25:44.834071Z",
          "score": 0,
          "model": "gpt-5-6-luna",
          "runner": "benchmark_agent",
          "config": "openai-gpt-5-6-luna-medium",
          "verified_publish": true,
          "levels_completed": 0,
          "actions": 140,
          "resets": 1,
          "state": "GAME_OVER",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            140,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            28,
            109,
            51,
            51,
            33,
            132,
            326
          ]
        },
        {
          "session_id": "0ad270cf-3d72-4839-8339-6657ae079363",
          "game_id": "ka59",
          "game_version": "38d34dbb",
          "published_at": "2026-06-24T03:38:22.257964Z",
          "score": 0,
          "model": "gpt-5-6-luna",
          "runner": "benchmark_agent",
          "config": "openai-gpt-5-6-luna-high",
          "verified_publish": true,
          "levels_completed": 0,
          "actions": 140,
          "resets": 21,
          "state": "NOT_FINISHED",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            140,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            28,
            109,
            51,
            51,
            33,
            132,
            326
          ]
        },
        {
          "session_id": "85e55fce-69a4-463a-bb46-9d52171d8190",
          "game_id": "ka59",
          "game_version": "38d34dbb",
          "published_at": "2026-06-24T00:14:17.967348Z",
          "score": 0,
          "model": "gpt-5-6-luna",
          "runner": "benchmark_agent",
          "config": "openai-gpt-5-6-luna-xhigh",
          "verified_publish": true,
          "levels_completed": 0,
          "actions": 140,
          "resets": 35,
          "state": "NOT_FINISHED",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            140,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            28,
            109,
            51,
            51,
            33,
            132,
            326
          ]
        },
        {
          "session_id": "2ba7c36a-f0f5-43ea-9547-23f0ee496c5d",
          "game_id": "ka59",
          "game_version": "38d34dbb",
          "published_at": "2026-06-23T01:08:22.601911Z",
          "score": 0,
          "model": "gpt-5-6-luna",
          "runner": "benchmark_agent",
          "config": "openai-gpt-5-6-luna-max",
          "verified_publish": true,
          "levels_completed": 0,
          "actions": 140,
          "resets": 59,
          "state": "NOT_FINISHED",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            140,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            28,
            109,
            51,
            51,
            33,
            132,
            326
          ]
        },
        {
          "session_id": "3bef25db-6d12-424b-88b5-1079080f504f",
          "game_id": "ka59",
          "game_version": "38d34dbb",
          "published_at": "2026-06-20T11:58:53.206503Z",
          "score": 0,
          "model": "gpt-5-6-terra",
          "runner": "benchmark_agent",
          "config": "openai-gpt-5-6-terra-low",
          "verified_publish": true,
          "levels_completed": 0,
          "actions": 140,
          "resets": 2,
          "state": "GAME_OVER",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            140,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            28,
            109,
            51,
            51,
            33,
            132,
            326
          ]
        },
        {
          "session_id": "6ccbee36-d0cf-4e6d-948c-27a1b31e52b2",
          "game_id": "ka59",
          "game_version": "38d34dbb",
          "published_at": "2026-06-20T09:15:53.392759Z",
          "score": 0,
          "model": "gpt-5-6-terra",
          "runner": "benchmark_agent",
          "config": "openai-gpt-5-6-terra-medium",
          "verified_publish": true,
          "levels_completed": 0,
          "actions": 140,
          "resets": 2,
          "state": "GAME_OVER",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            140,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            28,
            109,
            51,
            51,
            33,
            132,
            326
          ]
        },
        {
          "session_id": "2b24ac12-afab-4173-a9ac-3f9d46928eda",
          "game_id": "ka59",
          "game_version": "38d34dbb",
          "published_at": "2026-06-20T08:53:51.243467Z",
          "score": 0,
          "model": "gpt-5-6-terra",
          "runner": "benchmark_agent",
          "config": "openai-gpt-5-6-terra-high",
          "verified_publish": true,
          "levels_completed": 0,
          "actions": 140,
          "resets": 6,
          "state": "GAME_OVER",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            140,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            28,
            109,
            51,
            51,
            33,
            132,
            326
          ]
        },
        {
          "session_id": "b7644bcf-4d9f-4d57-8bdc-7a459bf23846",
          "game_id": "ka59",
          "game_version": "38d34dbb",
          "published_at": "2026-06-20T01:50:36.176556Z",
          "score": 0,
          "model": "gpt-5-6-terra",
          "runner": "benchmark_agent",
          "config": "openai-gpt-5-6-terra-xhigh",
          "verified_publish": true,
          "levels_completed": 0,
          "actions": 140,
          "resets": 5,
          "state": "NOT_FINISHED",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            140,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            28,
            109,
            51,
            51,
            33,
            132,
            326
          ]
        },
        {
          "session_id": "016d4317-b0eb-4844-86cd-1eeebf97c41c",
          "game_id": "ka59",
          "game_version": "38d34dbb",
          "published_at": "2026-06-19T07:21:30.816348Z",
          "score": 2.0452885317750185,
          "model": "gpt-5-6-terra",
          "runner": "benchmark_agent",
          "config": "openai-gpt-5-6-terra-max",
          "verified_publish": true,
          "levels_completed": 1,
          "actions": 582,
          "resets": 53,
          "state": "NOT_FINISHED",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            57.26807888970051,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            37,
            545,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            28,
            109,
            51,
            51,
            33,
            132,
            326
          ]
        },
        {
          "session_id": "81dbbb97-ac39-4bb4-ba8b-81373eb99e82",
          "game_id": "ka59",
          "game_version": "38d34dbb",
          "published_at": "2026-06-13T11:45:42.648385Z",
          "score": 3.571428571428571,
          "model": "gpt-5-6-sol",
          "runner": "benchmark_agent",
          "config": "openai-gpt-5-6-sol-max",
          "verified_publish": true,
          "levels_completed": 1,
          "actions": 562,
          "resets": 9,
          "state": "NOT_FINISHED",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            115,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            17,
            545,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            28,
            109,
            51,
            51,
            33,
            132,
            326
          ]
        },
        {
          "session_id": "92d4a687-c276-4e38-81b4-30e597f24aaa",
          "game_id": "ka59",
          "game_version": "38d34dbb",
          "published_at": "2026-06-11T23:00:54.287545Z",
          "score": 3.571428571428571,
          "model": "gpt-5-6-sol",
          "runner": "benchmark_agent",
          "config": "openai-gpt-5-6-sol-xhigh",
          "verified_publish": true,
          "levels_completed": 1,
          "actions": 566,
          "resets": 8,
          "state": "NOT_FINISHED",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            115,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            21,
            545,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            28,
            109,
            51,
            51,
            33,
            132,
            326
          ]
        },
        {
          "session_id": "c43091a2-8f92-46ed-a758-5afecb841cfc",
          "game_id": "ka59",
          "game_version": "38d34dbb",
          "published_at": "2026-06-10T16:16:00.144141Z",
          "score": 0.6237469369570059,
          "model": "gpt-5-6-sol",
          "runner": "benchmark_agent",
          "config": "openai-gpt-5-6-sol-high",
          "verified_publish": true,
          "levels_completed": 1,
          "actions": 612,
          "resets": 9,
          "state": "GAME_OVER",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            17.464914234796165,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            67,
            545,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            28,
            109,
            51,
            51,
            33,
            132,
            326
          ]
        },
        {
          "session_id": "cd9fc8ec-fbe5-42a8-ac1a-97d29550da49",
          "game_id": "ka59",
          "game_version": "38d34dbb",
          "published_at": "2026-06-10T10:41:36.609206Z",
          "score": 3.571428571428571,
          "model": "gpt-5-6-sol",
          "runner": "benchmark_agent",
          "config": "openai-gpt-5-6-sol-medium",
          "verified_publish": true,
          "levels_completed": 1,
          "actions": 564,
          "resets": 5,
          "state": "GAME_OVER",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            115,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            19,
            545,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            28,
            109,
            51,
            51,
            33,
            132,
            326
          ]
        },
        {
          "session_id": "1f2510f8-9dc2-4584-9a63-cf67c97483cf",
          "game_id": "ka59",
          "game_version": "38d34dbb",
          "published_at": "2026-06-09T22:20:29.766047Z",
          "score": 0.7054673721340388,
          "model": "gpt-5-6-sol",
          "runner": "benchmark_agent",
          "config": "openai-gpt-5-6-sol-low",
          "verified_publish": true,
          "levels_completed": 1,
          "actions": 608,
          "resets": 4,
          "state": "GAME_OVER",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            19.753086419753085,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            63,
            545,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            28,
            109,
            51,
            51,
            33,
            132,
            326
          ]
        },
        {
          "session_id": "40135f7f-87d2-4096-aa98-fbca36ded875",
          "game_id": "ka59",
          "game_version": "38d34dbb",
          "published_at": "2026-05-29T17:41:01.618532Z",
          "score": 0,
          "model": "claude-opus-4-8",
          "runner": "benchmark_agent",
          "config": "anthropic-opus-4-8-high",
          "verified_publish": true,
          "levels_completed": 0,
          "actions": 140,
          "resets": 21,
          "state": "NOT_FINISHED",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            140,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            28,
            109,
            51,
            51,
            33,
            132,
            326
          ]
        },
        {
          "session_id": "63d516d9-b787-44e3-8f6e-7fae12ebc98a",
          "game_id": "ka59",
          "game_version": "38d34dbb",
          "published_at": "2026-05-01T16:15:16.895124Z",
          "score": 0,
          "model": "gpt-5.5-2026-04-23",
          "runner": "benchmark_agent",
          "config": "openai-gpt-5-5-2026-04-23-high",
          "verified_publish": true,
          "levels_completed": 0,
          "actions": 140,
          "resets": 9,
          "state": "NOT_FINISHED",
          "completed": false,
          "number_of_levels": 7,
          "level_scores": [
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            140,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            28,
            109,
            51,
            51,
            33,
            132,
            326
          ]
        },
        {
          "session_id": "1dfbb619-2f6a-484a-89c8-bbe64abc977b",
          "game_id": "ka59",
          "game_version": "38d34dbb",
          "published_at": "2026-05-01T16:12:45.853961Z",
          "score": 2.0452885317750185,
          "model": "claude-opus-4-7",
          "runner": "benchmark_agent",
          "config": "anthropic-opus-4-7-high",
          "verified_publish": true,
          "levels_completed": 1,
          "actions": 582,
          "resets": 9,
          "state": "GAME_OVER",
          "completed": false,
          "number_of_levels": 7,
          "level_scores": [
            57.26807888970051,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            37,
            545,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            28,
            109,
            51,
            51,
            33,
            132,
            326
          ]
        },
        {
          "session_id": "b2fb7b5e-a475-40b6-8b23-a9a6ad49138d",
          "game_id": "ka59",
          "game_version": "38d34dbb",
          "published_at": "2026-04-14T17:44:01.163025Z",
          "score": 0,
          "model": "grok-4.20-beta-0309-reasoning",
          "runner": "benchmark_agent",
          "config": "xai-grok-4-20-beta-0309-reasoning",
          "verified_publish": true,
          "levels_completed": 0,
          "actions": 140,
          "resets": 3,
          "state": "NOT_FINISHED",
          "completed": false,
          "number_of_levels": 7,
          "level_scores": [
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            140,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            28,
            109,
            51,
            51,
            33,
            132,
            326
          ]
        },
        {
          "session_id": "5808cb69-c295-43d9-a1b8-dfc354d76951",
          "game_id": "ka59",
          "game_version": "38d34dbb",
          "published_at": "2026-04-14T17:44:00.897661Z",
          "score": 0,
          "model": "gpt-5.4-2026-03-05",
          "runner": "benchmark_agent",
          "config": "openai-gpt-5-4-2026-03-05-high",
          "verified_publish": true,
          "levels_completed": 0,
          "actions": 140,
          "resets": 1,
          "state": "NOT_FINISHED",
          "completed": false,
          "number_of_levels": 7,
          "level_scores": [
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            140,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            28,
            109,
            51,
            51,
            33,
            132,
            326
          ]
        },
        {
          "session_id": "6c0b23f8-63e3-450a-864e-14771ca902fd",
          "game_id": "ka59",
          "game_version": "38d34dbb",
          "published_at": "2026-04-14T17:44:00.22406Z",
          "score": 0,
          "model": "gemini-3.1-pro-preview",
          "runner": "benchmark_agent",
          "config": "google-gemini-3-1-pro-preview",
          "verified_publish": true,
          "levels_completed": 0,
          "actions": 43,
          "resets": 2,
          "state": "NOT_FINISHED",
          "completed": false,
          "number_of_levels": 7,
          "level_scores": [
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            43,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            28,
            109,
            51,
            51,
            33,
            132,
            326
          ]
        },
        {
          "session_id": "19fb0537-1395-45b5-9a0d-620ec669fbfd",
          "game_id": "ka59",
          "game_version": "38d34dbb",
          "published_at": "2026-04-14T17:43:59.782822Z",
          "score": 0,
          "model": "claude-opus-4-6",
          "runner": "benchmark_agent",
          "config": "anthropic-opus-4-6-max-effort",
          "verified_publish": true,
          "levels_completed": 0,
          "actions": 140,
          "resets": 12,
          "state": "NOT_FINISHED",
          "completed": false,
          "number_of_levels": 7,
          "level_scores": [
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            140,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            28,
            109,
            51,
            51,
            33,
            132,
            326
          ]
        }
      ]
    },
    {
      "game": "ar25",
      "source": "https://arcprize.org/tasks/ar25",
      "runs": [
        {
          "session_id": "df641dae-7cb4-4f91-ad5e-944fb8c4a5f3",
          "game_id": "ar25",
          "game_version": "0c556536",
          "published_at": "2026-09-25T15:24:35.573571Z",
          "score": 8.333333333333332,
          "model": "gpt-6-sol",
          "runner": "benchmark_agent",
          "config": "openai-gpt-6-sol-max",
          "verified_publish": true,
          "levels_completed": 2,
          "actions": 317,
          "resets": 1,
          "state": "GAME_OVER",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            115,
            115,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            20,
            38,
            259,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            32,
            50,
            75,
            37,
            89,
            159,
            233,
            73
          ]
        },
        {
          "session_id": "05b29877-e074-43b4-bb2e-d41a8c33ffb7",
          "game_id": "ar25",
          "game_version": "0c556536",
          "published_at": "2026-09-25T08:36:25.521393Z",
          "score": 6.02891156462585,
          "model": "gpt-6-sol",
          "runner": "benchmark_agent",
          "config": "openai-gpt-6-sol-xhigh",
          "verified_publish": true,
          "levels_completed": 2,
          "actions": 468,
          "resets": 2,
          "state": "GAME_OVER",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            115,
            51.02040816326531,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            23,
            70,
            375,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            32,
            50,
            75,
            37,
            89,
            159,
            233,
            73
          ]
        },
        {
          "session_id": "81e2d3ee-7f86-4b57-8002-f1411d2e99e3",
          "game_id": "ar25",
          "game_version": "0c556536",
          "published_at": "2026-09-25T07:17:07.174862Z",
          "score": 40.23145664152278,
          "model": "gpt-6-sol",
          "runner": "benchmark_agent",
          "config": "openai-gpt-6-sol-xhigh-provider-adapter",
          "verified_publish": true,
          "levels_completed": 5,
          "actions": 1041,
          "resets": 1,
          "state": "GAME_OVER",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            115,
            115,
            65.03642039542143,
            115,
            89.64463558171117,
            0,
            0,
            0
          ],
          "level_actions": [
            19,
            18,
            93,
            22,
            94,
            795,
            0,
            0
          ],
          "level_baseline_actions": [
            32,
            50,
            75,
            37,
            89,
            159,
            233,
            73
          ]
        },
        {
          "session_id": "6a6c2d28-c422-4c0e-a7b4-14f1932b97d3",
          "game_id": "ar25",
          "game_version": "0c556536",
          "published_at": "2026-09-25T07:05:37.128928Z",
          "score": 1.2798955658399631,
          "model": "gpt-6-sol",
          "runner": "benchmark_agent",
          "config": "openai-gpt-6-sol-none-provider-adapter",
          "verified_publish": true,
          "levels_completed": 2,
          "actions": 624,
          "resets": 2,
          "state": "GAME_OVER",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            32.6530612244898,
            6.711589572874438,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            56,
            193,
            375,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            32,
            50,
            75,
            37,
            89,
            159,
            233,
            73
          ]
        },
        {
          "session_id": "3f248555-1eba-4a8e-831a-11fb17a3f720",
          "game_id": "ar25",
          "game_version": "0c556536",
          "published_at": "2026-09-25T06:25:36.515622Z",
          "score": 0.11394185404760633,
          "model": "gpt-6-sol",
          "runner": "benchmark_agent",
          "config": "openai-gpt-6-sol-none",
          "verified_publish": true,
          "levels_completed": 1,
          "actions": 251,
          "resets": 4,
          "state": "GAME_OVER",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            4.101906745713828,
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            158,
            93,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            32,
            50,
            75,
            37,
            89,
            159,
            233,
            73
          ]
        },
        {
          "session_id": "db4c7440-1d88-4bb1-b526-d55e69ae32c3",
          "game_id": "ar25",
          "game_version": "0c556536",
          "published_at": "2026-09-25T06:17:33.513573Z",
          "score": 2.7777777777777777,
          "model": "gpt-6-sol",
          "runner": "benchmark_agent",
          "config": "openai-gpt-6-sol-low-provider-adapter",
          "verified_publish": true,
          "levels_completed": 1,
          "actions": 273,
          "resets": 2,
          "state": "GAME_OVER",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            115,
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            23,
            250,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            32,
            50,
            75,
            37,
            89,
            159,
            233,
            73
          ]
        },
        {
          "session_id": "2140d6e7-4349-4e62-8344-37b3ea5e775a",
          "game_id": "ar25",
          "game_version": "0c556536",
          "published_at": "2026-09-25T05:57:01.553642Z",
          "score": 0,
          "model": "gpt-6-sol",
          "runner": "benchmark_agent",
          "config": "openai-gpt-6-sol-low",
          "verified_publish": true,
          "levels_completed": 0,
          "actions": 160,
          "resets": 2,
          "state": "GAME_OVER",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            0,
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            160,
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            32,
            50,
            75,
            37,
            89,
            159,
            233,
            73
          ]
        },
        {
          "session_id": "c05ad303-f3d7-4640-9280-be8a6edae768",
          "game_id": "ar25",
          "game_version": "0c556536",
          "published_at": "2026-09-25T05:50:46.092287Z",
          "score": 2.7777777777777777,
          "model": "gpt-6-sol",
          "runner": "benchmark_agent",
          "config": "openai-gpt-6-sol-medium-provider-adapter",
          "verified_publish": true,
          "levels_completed": 1,
          "actions": 271,
          "resets": 2,
          "state": "GAME_OVER",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            115,
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            21,
            250,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            32,
            50,
            75,
            37,
            89,
            159,
            233,
            73
          ]
        },
        {
          "session_id": "1a7045ba-6835-4d90-af13-43cc70435098",
          "game_id": "ar25",
          "game_version": "0c556536",
          "published_at": "2026-09-25T05:33:20.893306Z",
          "score": 4.158950617283951,
          "model": "gpt-6-sol",
          "runner": "benchmark_agent",
          "config": "openai-gpt-6-sol-high-provider-adapter",
          "verified_publish": true,
          "levels_completed": 2,
          "actions": 517,
          "resets": 3,
          "state": "GAME_OVER",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            115,
            17.361111111111114,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            22,
            120,
            375,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            32,
            50,
            75,
            37,
            89,
            159,
            233,
            73
          ]
        },
        {
          "session_id": "7b7e18cf-99e8-4f69-a671-74b5f7923dce",
          "game_id": "ar25",
          "game_version": "0c556536",
          "published_at": "2026-09-25T05:20:06.169983Z",
          "score": 0,
          "model": "gpt-6-sol",
          "runner": "benchmark_agent",
          "config": "openai-gpt-6-sol-medium",
          "verified_publish": true,
          "levels_completed": 0,
          "actions": 160,
          "resets": 4,
          "state": "GAME_OVER",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            0,
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            160,
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            32,
            50,
            75,
            37,
            89,
            159,
            233,
            73
          ]
        },
        {
          "session_id": "dd522261-689d-409b-8480-9e2372566cb1",
          "game_id": "ar25",
          "game_version": "0c556536",
          "published_at": "2026-09-25T04:48:06.682775Z",
          "score": 0,
          "model": "gpt-6-sol",
          "runner": "benchmark_agent",
          "config": "openai-gpt-6-sol-high",
          "verified_publish": true,
          "levels_completed": 0,
          "actions": 160,
          "resets": 1,
          "state": "GAME_OVER",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            0,
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            160,
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            32,
            50,
            75,
            37,
            89,
            159,
            233,
            73
          ]
        },
        {
          "session_id": "cd438f66-55ad-4f24-a529-9785acd1063d",
          "game_id": "ar25",
          "game_version": "0c556536",
          "published_at": "2026-09-21T09:18:51.007355Z",
          "score": 58.333333333333336,
          "model": "gemini-3.8-flash",
          "runner": "benchmark_agent",
          "config": "google-gemini-3-8-flash-low-provider-adapter",
          "verified_publish": true,
          "levels_completed": 6,
          "actions": 1003,
          "resets": 2,
          "state": "GAME_OVER",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            115,
            115,
            115,
            115,
            115,
            115,
            0,
            0
          ],
          "level_actions": [
            19,
            25,
            53,
            30,
            40,
            61,
            775,
            0
          ],
          "level_baseline_actions": [
            32,
            50,
            75,
            37,
            89,
            159,
            233,
            73
          ]
        },
        {
          "session_id": "1ba4006a-d5be-435b-81e8-bc90c7f82ded",
          "game_id": "ar25",
          "game_version": "0c556536",
          "published_at": "2026-09-19T17:08:49.32257Z",
          "score": 100,
          "model": "gemini-3.8-flash",
          "runner": "benchmark_agent",
          "config": "google-gemini-3-8-flash-medium-provider-adapter",
          "verified_publish": true,
          "levels_completed": 8,
          "actions": 340,
          "resets": 0,
          "state": "WIN",
          "completed": true,
          "number_of_levels": 0,
          "level_scores": [
            115,
            115,
            115,
            115,
            115,
            115,
            115,
            115
          ],
          "level_actions": [
            17,
            13,
            43,
            22,
            37,
            61,
            90,
            57
          ],
          "level_baseline_actions": [
            32,
            50,
            75,
            37,
            89,
            159,
            233,
            73
          ]
        },
        {
          "session_id": "a7329ff9-1763-463a-9349-7077492b0de0",
          "game_id": "ar25",
          "game_version": "0c556536",
          "published_at": "2026-09-17T14:55:02.092897Z",
          "score": 100,
          "model": "gemini-3.8-flash",
          "runner": "benchmark_agent",
          "config": "google-gemini-3-8-flash-high-provider-adapter",
          "verified_publish": true,
          "levels_completed": 8,
          "actions": 275,
          "resets": 0,
          "state": "WIN",
          "completed": true,
          "number_of_levels": 0,
          "level_scores": [
            115,
            115,
            115,
            115,
            115,
            115,
            115,
            115
          ],
          "level_actions": [
            17,
            12,
            41,
            22,
            31,
            53,
            52,
            47
          ],
          "level_baseline_actions": [
            32,
            50,
            75,
            37,
            89,
            159,
            233,
            73
          ]
        },
        {
          "session_id": "90838560-9e26-4af3-acd8-cf160a95587d",
          "game_id": "ar25",
          "game_version": "0c556536",
          "published_at": "2026-09-16T20:24:54.950466Z",
          "score": 0,
          "model": "gpt-6-luna",
          "runner": "benchmark_agent",
          "config": "openai-gpt-6-luna-low",
          "verified_publish": true,
          "levels_completed": 0,
          "actions": 160,
          "resets": 3,
          "state": "NOT_FINISHED",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            0,
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            160,
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            32,
            50,
            75,
            37,
            89,
            159,
            233,
            73
          ]
        },
        {
          "session_id": "00aa6eef-ac1e-4767-b325-3ec2f8e8736f",
          "game_id": "ar25",
          "game_version": "0c556536",
          "published_at": "2026-09-16T19:51:01.524689Z",
          "score": 23.536419930415907,
          "model": "gpt-6-1-sol",
          "runner": "benchmark_agent",
          "config": "openai-gpt-6-1-sol-low",
          "verified_publish": true,
          "levels_completed": 4,
          "actions": 588,
          "resets": 3,
          "state": "GAME_OVER",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            115,
            115,
            115,
            39.32777937374318,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            22,
            13,
            49,
            59,
            445,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            32,
            50,
            75,
            37,
            89,
            159,
            233,
            73
          ]
        },
        {
          "session_id": "ef813b79-6266-4686-92be-24da2f51e8a6",
          "game_id": "ar25",
          "game_version": "0c556536",
          "published_at": "2026-09-16T19:37:03.819451Z",
          "score": 100,
          "model": "gpt-6-1-sol",
          "runner": "benchmark_agent",
          "config": "openai-gpt-6-1-sol-medium",
          "verified_publish": true,
          "levels_completed": 8,
          "actions": 323,
          "resets": 0,
          "state": "WIN",
          "completed": true,
          "number_of_levels": 0,
          "level_scores": [
            115,
            115,
            115,
            115,
            115,
            115,
            115,
            105.71315215235072
          ],
          "level_actions": [
            21,
            13,
            56,
            23,
            30,
            61,
            48,
            71
          ],
          "level_baseline_actions": [
            32,
            50,
            75,
            37,
            89,
            159,
            233,
            73
          ]
        },
        {
          "session_id": "f090e130-4832-42c8-be5b-ba6991c5c48a",
          "game_id": "ar25",
          "game_version": "0c556536",
          "published_at": "2026-09-16T19:26:19.391592Z",
          "score": 0,
          "model": "gpt-6-luna",
          "runner": "benchmark_agent",
          "config": "openai-gpt-6-luna-low-provider-adapter",
          "verified_publish": true,
          "levels_completed": 0,
          "actions": 160,
          "resets": 1,
          "state": "GAME_OVER",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            0,
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            160,
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            32,
            50,
            75,
            37,
            89,
            159,
            233,
            73
          ]
        },
        {
          "session_id": "d0a9f801-257a-44ea-bd11-db22f53926f6",
          "game_id": "ar25",
          "game_version": "0c556536",
          "published_at": "2026-09-16T19:25:20.09848Z",
          "score": 0,
          "model": "gpt-6-luna",
          "runner": "benchmark_agent",
          "config": "openai-gpt-6-luna-none-provider-adapter",
          "verified_publish": true,
          "levels_completed": 0,
          "actions": 160,
          "resets": 5,
          "state": "NOT_FINISHED",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            0,
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            160,
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            32,
            50,
            75,
            37,
            89,
            159,
            233,
            73
          ]
        },
        {
          "session_id": "356730ca-8598-4072-ba09-4ae2528a525c",
          "game_id": "ar25",
          "game_version": "0c556536",
          "published_at": "2026-09-16T19:24:13.292229Z",
          "score": 0,
          "model": "gpt-6-luna",
          "runner": "benchmark_agent",
          "config": "openai-gpt-6-luna-none",
          "verified_publish": true,
          "levels_completed": 0,
          "actions": 160,
          "resets": 2,
          "state": "NOT_FINISHED",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            0,
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            160,
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            32,
            50,
            75,
            37,
            89,
            159,
            233,
            73
          ]
        },
        {
          "session_id": "9c4b15d4-02bb-4503-9406-17d68ccb0222",
          "game_id": "ar25",
          "game_version": "0c556536",
          "published_at": "2026-09-16T15:45:34.383697Z",
          "score": 0,
          "model": "gpt-6-luna",
          "runner": "benchmark_agent",
          "config": "openai-gpt-6-luna-high-provider-adapter",
          "verified_publish": true,
          "levels_completed": 0,
          "actions": 160,
          "resets": 2,
          "state": "GAME_OVER",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            0,
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            160,
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            32,
            50,
            75,
            37,
            89,
            159,
            233,
            73
          ]
        },
        {
          "session_id": "a7d55320-a119-4b4c-bac8-7d96c81cfe3b",
          "game_id": "ar25",
          "game_version": "0c556536",
          "published_at": "2026-09-16T09:53:58.171668Z",
          "score": 1.4046639231824418,
          "model": "gpt-6-luna",
          "runner": "benchmark_agent",
          "config": "openai-gpt-6-luna-medium-provider-adapter",
          "verified_publish": true,
          "levels_completed": 1,
          "actions": 192,
          "resets": 1,
          "state": "GAME_OVER",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            50.567901234567906,
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            45,
            147,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            32,
            50,
            75,
            37,
            89,
            159,
            233,
            73
          ]
        },
        {
          "session_id": "e575120b-f41c-4f39-83b8-c110844e1ba9",
          "game_id": "ar25",
          "game_version": "0c556536",
          "published_at": "2026-09-16T08:57:30.002606Z",
          "score": 0,
          "model": "gpt-6-luna",
          "runner": "benchmark_agent",
          "config": "openai-gpt-6-luna-medium",
          "verified_publish": true,
          "levels_completed": 0,
          "actions": 160,
          "resets": 2,
          "state": "NOT_FINISHED",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            0,
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            160,
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            32,
            50,
            75,
            37,
            89,
            159,
            233,
            73
          ]
        },
        {
          "session_id": "ade046e6-416f-4446-a8c7-4dae789a5738",
          "game_id": "ar25",
          "game_version": "0c556536",
          "published_at": "2026-09-16T08:42:16.659468Z",
          "score": 0,
          "model": "gpt-6-luna",
          "runner": "benchmark_agent",
          "config": "openai-gpt-6-luna-high",
          "verified_publish": true,
          "levels_completed": 0,
          "actions": 160,
          "resets": 1,
          "state": "GAME_OVER",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            0,
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            160,
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            32,
            50,
            75,
            37,
            89,
            159,
            233,
            73
          ]
        },
        {
          "session_id": "3b235b24-cfe8-444a-bc4e-781893787f9b",
          "game_id": "ar25",
          "game_version": "0c556536",
          "published_at": "2026-09-16T08:22:56.160913Z",
          "score": 0.302311026086135,
          "model": "gpt-6-luna",
          "runner": "benchmark_agent",
          "config": "openai-gpt-6-luna-xhigh-provider-adapter",
          "verified_publish": true,
          "levels_completed": 1,
          "actions": 347,
          "resets": 4,
          "state": "GAME_OVER",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            10.883196939100861,
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            97,
            250,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            32,
            50,
            75,
            37,
            89,
            159,
            233,
            73
          ]
        },
        {
          "session_id": "8f0425d6-5462-498e-a762-4e3eff2b4801",
          "game_id": "ar25",
          "game_version": "0c556536",
          "published_at": "2026-09-16T07:54:40.778812Z",
          "score": 0,
          "model": "gpt-6-luna",
          "runner": "benchmark_agent",
          "config": "openai-gpt-6-luna-max",
          "verified_publish": true,
          "levels_completed": 0,
          "actions": 160,
          "resets": 9,
          "state": "NOT_FINISHED",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            0,
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            160,
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            32,
            50,
            75,
            37,
            89,
            159,
            233,
            73
          ]
        },
        {
          "session_id": "5cd769e2-f1b8-44fe-9596-62864f30ca29",
          "game_id": "ar25",
          "game_version": "0c556536",
          "published_at": "2026-09-16T07:47:10.727598Z",
          "score": 0,
          "model": "gpt-6-luna",
          "runner": "benchmark_agent",
          "config": "openai-gpt-6-luna-xhigh",
          "verified_publish": true,
          "levels_completed": 0,
          "actions": 160,
          "resets": 3,
          "state": "GAME_OVER",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            0,
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            160,
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            32,
            50,
            75,
            37,
            89,
            159,
            233,
            73
          ]
        },
        {
          "session_id": "7f19e6b6-abe8-43d6-b608-6b6ea75a1567",
          "game_id": "ar25",
          "game_version": "0c556536",
          "published_at": "2026-09-16T06:10:31.270388Z",
          "score": 100,
          "model": "gpt-6-1-sol",
          "runner": "benchmark_agent",
          "config": "openai-gpt-6-1-sol-max",
          "verified_publish": true,
          "levels_completed": 8,
          "actions": 281,
          "resets": 0,
          "state": "WIN",
          "completed": true,
          "number_of_levels": 0,
          "level_scores": [
            115,
            115,
            115,
            115,
            115,
            115,
            115,
            115
          ],
          "level_actions": [
            17,
            11,
            40,
            22,
            30,
            68,
            43,
            50
          ],
          "level_baseline_actions": [
            32,
            50,
            75,
            37,
            89,
            159,
            233,
            73
          ]
        },
        {
          "session_id": "f879f959-ab5f-4a83-9f4b-c97dcc06a1db",
          "game_id": "ar25",
          "game_version": "0c556536",
          "published_at": "2026-09-16T06:10:25.613721Z",
          "score": 100,
          "model": "gpt-6-1-sol",
          "runner": "benchmark_agent",
          "config": "openai-gpt-6-1-sol-high",
          "verified_publish": true,
          "levels_completed": 8,
          "actions": 434,
          "resets": 6,
          "state": "WIN",
          "completed": true,
          "number_of_levels": 0,
          "level_scores": [
            115,
            115,
            115,
            115,
            115,
            115,
            115,
            115
          ],
          "level_actions": [
            18,
            13,
            40,
            23,
            29,
            57,
            206,
            48
          ],
          "level_baseline_actions": [
            32,
            50,
            75,
            37,
            89,
            159,
            233,
            73
          ]
        },
        {
          "session_id": "3b571eb6-b82d-4697-a0eb-f34aeed902a6",
          "game_id": "ar25",
          "game_version": "0c556536",
          "published_at": "2026-09-16T04:24:32.761584Z",
          "score": 100,
          "model": "gpt-6-1-sol",
          "runner": "benchmark_agent",
          "config": "openai-gpt-6-1-sol-xhigh",
          "verified_publish": true,
          "levels_completed": 8,
          "actions": 267,
          "resets": 0,
          "state": "WIN",
          "completed": true,
          "number_of_levels": 0,
          "level_scores": [
            115,
            115,
            115,
            115,
            115,
            115,
            115,
            115
          ],
          "level_actions": [
            18,
            14,
            40,
            23,
            32,
            54,
            37,
            49
          ],
          "level_baseline_actions": [
            32,
            50,
            75,
            37,
            89,
            159,
            233,
            73
          ]
        },
        {
          "session_id": "4259c787-104e-4cbf-94e4-1d2b2ece713c",
          "game_id": "ar25",
          "game_version": "0c556536",
          "published_at": "2026-09-16T04:19:44.871827Z",
          "score": 100,
          "model": "gpt-6-1-sol",
          "runner": "benchmark_agent",
          "config": "openai-gpt-6-1-sol-high-provider-adapter",
          "verified_publish": true,
          "levels_completed": 8,
          "actions": 272,
          "resets": 0,
          "state": "WIN",
          "completed": true,
          "number_of_levels": 0,
          "level_scores": [
            115,
            115,
            115,
            115,
            115,
            115,
            115,
            115
          ],
          "level_actions": [
            21,
            13,
            40,
            22,
            30,
            53,
            46,
            47
          ],
          "level_baseline_actions": [
            32,
            50,
            75,
            37,
            89,
            159,
            233,
            73
          ]
        },
        {
          "session_id": "5e4ec6a3-c77e-440b-ad4f-6c208e9cb9f3",
          "game_id": "ar25",
          "game_version": "0c556536",
          "published_at": "2026-09-16T03:40:14.271463Z",
          "score": 100,
          "model": "gpt-6-1-sol",
          "runner": "benchmark_agent",
          "config": "openai-gpt-6-1-sol-max-provider-adapter",
          "verified_publish": true,
          "levels_completed": 8,
          "actions": 259,
          "resets": 0,
          "state": "WIN",
          "completed": true,
          "number_of_levels": 0,
          "level_scores": [
            115,
            115,
            115,
            115,
            115,
            115,
            115,
            115
          ],
          "level_actions": [
            15,
            11,
            40,
            27,
            29,
            53,
            37,
            47
          ],
          "level_baseline_actions": [
            32,
            50,
            75,
            37,
            89,
            159,
            233,
            73
          ]
        },
        {
          "session_id": "479fc76c-3b84-4262-94cc-9f476c576a70",
          "game_id": "ar25",
          "game_version": "0c556536",
          "published_at": "2026-09-16T03:30:34.985166Z",
          "score": 100,
          "model": "gpt-6-1-sol",
          "runner": "benchmark_agent",
          "config": "openai-gpt-6-1-sol-xhigh-provider-adapter",
          "verified_publish": true,
          "levels_completed": 8,
          "actions": 256,
          "resets": 0,
          "state": "WIN",
          "completed": true,
          "number_of_levels": 0,
          "level_scores": [
            115,
            115,
            115,
            115,
            115,
            115,
            115,
            115
          ],
          "level_actions": [
            16,
            11,
            40,
            22,
            28,
            53,
            39,
            47
          ],
          "level_baseline_actions": [
            32,
            50,
            75,
            37,
            89,
            159,
            233,
            73
          ]
        },
        {
          "session_id": "bdac8bc2-8ee2-4854-81eb-a76b7ea3af69",
          "game_id": "ar25",
          "game_version": "0c556536",
          "published_at": "2026-09-15T22:47:00.930124Z",
          "score": 0.5974468482344979,
          "model": "gpt-6-luna",
          "runner": "benchmark_agent",
          "config": "openai-gpt-6-luna-max-provider-adapter",
          "verified_publish": true,
          "levels_completed": 1,
          "actions": 319,
          "resets": 6,
          "state": "GAME_OVER",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            21.508086536441926,
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            69,
            250,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            32,
            50,
            75,
            37,
            89,
            159,
            233,
            73
          ]
        },
        {
          "session_id": "34997caf-9008-41d2-8db9-46152a1d9236",
          "game_id": "ar25",
          "game_version": "0c556536",
          "published_at": "2026-09-12T15:40:58.497456Z",
          "score": 45.94954700276497,
          "model": "gemini-3.8-flash",
          "runner": "benchmark_agent",
          "config": "google-gemini-3-8-flash-high",
          "verified_publish": true,
          "levels_completed": 6,
          "actions": 717,
          "resets": 8,
          "state": "NOT_FINISHED",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            115,
            115,
            4.340277777777779,
            115,
            115,
            43.52714312770097,
            0,
            0
          ],
          "level_actions": [
            19,
            13,
            360,
            24,
            56,
            241,
            4,
            0
          ],
          "level_baseline_actions": [
            32,
            50,
            75,
            37,
            89,
            159,
            233,
            73
          ]
        },
        {
          "session_id": "c23a72be-a2c8-4580-948f-107e66f1fcf8",
          "game_id": "ar25",
          "game_version": "0c556536",
          "published_at": "2026-09-10T21:03:44.828493Z",
          "score": 12.337952222747903,
          "model": "gemini-3.8-flash",
          "runner": "benchmark_agent",
          "config": "google-gemini-3-8-flash-medium",
          "verified_publish": true,
          "levels_completed": 4,
          "actions": 798,
          "resets": 16,
          "state": "NOT_FINISHED",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            115,
            115,
            11.013000234944007,
            16.531819828523126,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            24,
            12,
            226,
            91,
            445,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            32,
            50,
            75,
            37,
            89,
            159,
            233,
            73
          ]
        },
        {
          "session_id": "903eb0b7-b3bb-4285-befe-baefc8ddc951",
          "game_id": "ar25",
          "game_version": "0c556536",
          "published_at": "2026-09-09T17:27:32.801762Z",
          "score": 16.666666666666664,
          "model": "gemini-3.8-flash",
          "runner": "benchmark_agent",
          "config": "google-gemini-3-8-flash-low",
          "verified_publish": true,
          "levels_completed": 3,
          "actions": 296,
          "resets": 1,
          "state": "GAME_OVER",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            115,
            115,
            92.45562130177515,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            19,
            14,
            78,
            185,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            32,
            50,
            75,
            37,
            89,
            159,
            233,
            73
          ]
        },
        {
          "session_id": "80240582-0d83-4183-a8c5-268b2d126cda",
          "game_id": "ar25",
          "game_version": "0c556536",
          "published_at": "2026-08-11T13:30:32.336126Z",
          "score": 8.333333333333332,
          "model": "grok-4.6",
          "runner": "benchmark_agent",
          "config": "xai-grok-4-6-xhigh",
          "verified_publish": true,
          "levels_completed": 2,
          "actions": 414,
          "resets": 1,
          "state": "GAME_OVER",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            115,
            115,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            21,
            18,
            375,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            32,
            50,
            75,
            37,
            89,
            159,
            233,
            73
          ]
        },
        {
          "session_id": "b71b5791-4826-4c49-ac6f-a75ff806ada1",
          "game_id": "ar25",
          "game_version": "0c556536",
          "published_at": "2026-07-21T02:32:51.914959Z",
          "score": 100,
          "model": "claude-opus-5",
          "runner": "benchmark_agent",
          "config": "anthropic-claude-opus-5-high",
          "verified_publish": true,
          "levels_completed": 8,
          "actions": 294,
          "resets": 0,
          "state": "WIN",
          "completed": true,
          "number_of_levels": 0,
          "level_scores": [
            115,
            115,
            115,
            115,
            115,
            115,
            115,
            115
          ],
          "level_actions": [
            18,
            18,
            44,
            23,
            45,
            53,
            46,
            47
          ],
          "level_baseline_actions": [
            32,
            50,
            75,
            37,
            89,
            159,
            233,
            73
          ]
        },
        {
          "session_id": "b53e7712-dd39-42db-aaa6-75706552a336",
          "game_id": "ar25",
          "game_version": "0c556536",
          "published_at": "2026-06-24T10:41:54.019503Z",
          "score": 0,
          "model": "gpt-5-6-luna",
          "runner": "benchmark_agent",
          "config": "openai-gpt-5-6-luna-low",
          "verified_publish": true,
          "levels_completed": 0,
          "actions": 160,
          "resets": 15,
          "state": "NOT_FINISHED",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            0,
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            160,
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            32,
            50,
            75,
            37,
            89,
            159,
            233,
            73
          ]
        },
        {
          "session_id": "f319ee9b-03b1-46dc-bc65-136815532eb7",
          "game_id": "ar25",
          "game_version": "0c556536",
          "published_at": "2026-06-24T10:15:44.943886Z",
          "score": 0,
          "model": "gpt-5-6-luna",
          "runner": "benchmark_agent",
          "config": "openai-gpt-5-6-luna-medium",
          "verified_publish": true,
          "levels_completed": 0,
          "actions": 160,
          "resets": 6,
          "state": "NOT_FINISHED",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            0,
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            160,
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            32,
            50,
            75,
            37,
            89,
            159,
            233,
            73
          ]
        },
        {
          "session_id": "314fe7b5-650c-4659-950c-624ebec60b96",
          "game_id": "ar25",
          "game_version": "0c556536",
          "published_at": "2026-06-24T02:47:16.691527Z",
          "score": 0,
          "model": "gpt-5-6-luna",
          "runner": "benchmark_agent",
          "config": "openai-gpt-5-6-luna-high",
          "verified_publish": true,
          "levels_completed": 0,
          "actions": 160,
          "resets": 18,
          "state": "NOT_FINISHED",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            0,
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            160,
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            32,
            50,
            75,
            37,
            89,
            159,
            233,
            73
          ]
        },
        {
          "session_id": "091add0a-8c43-42d4-b7aa-1da44c28e89b",
          "game_id": "ar25",
          "game_version": "0c556536",
          "published_at": "2026-06-23T23:05:38.189967Z",
          "score": 0,
          "model": "gpt-5-6-luna",
          "runner": "benchmark_agent",
          "config": "openai-gpt-5-6-luna-xhigh",
          "verified_publish": true,
          "levels_completed": 0,
          "actions": 160,
          "resets": 43,
          "state": "NOT_FINISHED",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            0,
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            160,
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            32,
            50,
            75,
            37,
            89,
            159,
            233,
            73
          ]
        },
        {
          "session_id": "9f93044e-eb80-4dd8-8429-e0a5aee97771",
          "game_id": "ar25",
          "game_version": "0c556536",
          "published_at": "2026-06-23T03:00:05.08863Z",
          "score": 0,
          "model": "gpt-5-6-luna",
          "runner": "benchmark_agent",
          "config": "openai-gpt-5-6-luna-max",
          "verified_publish": true,
          "levels_completed": 0,
          "actions": 160,
          "resets": 52,
          "state": "NOT_FINISHED",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            0,
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            160,
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            32,
            50,
            75,
            37,
            89,
            159,
            233,
            73
          ]
        },
        {
          "session_id": "efaa93e4-6d9e-4476-abdd-3ec360d709ae",
          "game_id": "ar25",
          "game_version": "0c556536",
          "published_at": "2026-06-20T11:52:05.676457Z",
          "score": 0,
          "model": "gpt-5-6-terra",
          "runner": "benchmark_agent",
          "config": "openai-gpt-5-6-terra-low",
          "verified_publish": true,
          "levels_completed": 0,
          "actions": 160,
          "resets": 1,
          "state": "GAME_OVER",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            0,
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            160,
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            32,
            50,
            75,
            37,
            89,
            159,
            233,
            73
          ]
        },
        {
          "session_id": "e69b5fab-159b-442e-a86d-17a479c62440",
          "game_id": "ar25",
          "game_version": "0c556536",
          "published_at": "2026-06-20T09:01:44.230343Z",
          "score": 0,
          "model": "gpt-5-6-terra",
          "runner": "benchmark_agent",
          "config": "openai-gpt-5-6-terra-medium",
          "verified_publish": true,
          "levels_completed": 0,
          "actions": 160,
          "resets": 1,
          "state": "GAME_OVER",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            0,
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            160,
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            32,
            50,
            75,
            37,
            89,
            159,
            233,
            73
          ]
        },
        {
          "session_id": "751e2518-0885-4b23-9212-ad325c1f06a0",
          "game_id": "ar25",
          "game_version": "0c556536",
          "published_at": "2026-06-20T08:04:46.15439Z",
          "score": 0,
          "model": "gpt-5-6-terra",
          "runner": "benchmark_agent",
          "config": "openai-gpt-5-6-terra-high",
          "verified_publish": true,
          "levels_completed": 0,
          "actions": 160,
          "resets": 3,
          "state": "GAME_OVER",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            0,
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            160,
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            32,
            50,
            75,
            37,
            89,
            159,
            233,
            73
          ]
        },
        {
          "session_id": "0ff1b834-0a50-483d-8932-dcb2eafdcab1",
          "game_id": "ar25",
          "game_version": "0c556536",
          "published_at": "2026-06-20T00:54:07.099895Z",
          "score": 0,
          "model": "gpt-5-6-terra",
          "runner": "benchmark_agent",
          "config": "openai-gpt-5-6-terra-xhigh",
          "verified_publish": true,
          "levels_completed": 0,
          "actions": 160,
          "resets": 2,
          "state": "GAME_OVER",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            0,
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            160,
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            32,
            50,
            75,
            37,
            89,
            159,
            233,
            73
          ]
        },
        {
          "session_id": "f29f6f64-2cd2-4d74-9cc5-9321928fa78d",
          "game_id": "ar25",
          "game_version": "0c556536",
          "published_at": "2026-06-19T07:07:30.780467Z",
          "score": 6.673333333333334,
          "model": "gpt-5-6-terra",
          "runner": "benchmark_agent",
          "config": "openai-gpt-5-6-terra-max",
          "verified_publish": true,
          "levels_completed": 2,
          "actions": 492,
          "resets": 48,
          "state": "NOT_FINISHED",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            10.24,
            115,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            100,
            17,
            375,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            32,
            50,
            75,
            37,
            89,
            159,
            233,
            73
          ]
        },
        {
          "session_id": "f018164b-9ca3-4983-aa20-44e71b2028ad",
          "game_id": "ar25",
          "game_version": "0c556536",
          "published_at": "2026-06-13T22:54:19.099002Z",
          "score": 38.824268702673834,
          "model": "gpt-5-6-sol",
          "runner": "benchmark_agent",
          "config": "openai-gpt-5-6-sol-max",
          "verified_publish": true,
          "levels_completed": 5,
          "actions": 680,
          "resets": 7,
          "state": "NOT_FINISHED",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            115,
            115,
            5.891224432085965,
            115,
            115,
            0,
            0,
            0
          ],
          "level_actions": [
            17,
            13,
            309,
            32,
            49,
            260,
            0,
            0
          ],
          "level_baseline_actions": [
            32,
            50,
            75,
            37,
            89,
            159,
            233,
            73
          ]
        },
        {
          "session_id": "1856d72e-b9d1-4d1d-af91-b18b09f6271a",
          "game_id": "ar25",
          "game_version": "0c556536",
          "published_at": "2026-06-11T18:40:12.896929Z",
          "score": 8.333333333333332,
          "model": "gpt-5-6-sol",
          "runner": "benchmark_agent",
          "config": "openai-gpt-5-6-sol-xhigh",
          "verified_publish": true,
          "levels_completed": 2,
          "actions": 410,
          "resets": 7,
          "state": "NOT_FINISHED",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            115,
            115,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            22,
            13,
            375,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            32,
            50,
            75,
            37,
            89,
            159,
            233,
            73
          ]
        },
        {
          "session_id": "8df6d14d-f20d-4f11-a553-58f923be9e45",
          "game_id": "ar25",
          "game_version": "0c556536",
          "published_at": "2026-06-10T14:33:59.979793Z",
          "score": 5.5369792545117225,
          "model": "gpt-5-6-sol",
          "runner": "benchmark_agent",
          "config": "openai-gpt-5-6-sol-high",
          "verified_publish": true,
          "levels_completed": 2,
          "actions": 471,
          "resets": 7,
          "state": "NOT_FINISHED",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            115,
            42.165626581210994,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            19,
            77,
            375,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            32,
            50,
            75,
            37,
            89,
            159,
            233,
            73
          ]
        },
        {
          "session_id": "14970c34-ab24-40a3-8fac-d3275ee6b61f",
          "game_id": "ar25",
          "game_version": "0c556536",
          "published_at": "2026-06-10T08:41:32.98763Z",
          "score": 8.333333333333332,
          "model": "gpt-5-6-sol",
          "runner": "benchmark_agent",
          "config": "openai-gpt-5-6-sol-medium",
          "verified_publish": true,
          "levels_completed": 2,
          "actions": 407,
          "resets": 3,
          "state": "GAME_OVER",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            115,
            115,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            17,
            15,
            375,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            32,
            50,
            75,
            37,
            89,
            159,
            233,
            73
          ]
        },
        {
          "session_id": "cb9852fa-5fff-4947-8e7b-144834789107",
          "game_id": "ar25",
          "game_version": "0c556536",
          "published_at": "2026-06-09T21:51:41.349628Z",
          "score": 0.8754830546150951,
          "model": "gpt-5-6-sol",
          "runner": "benchmark_agent",
          "config": "openai-gpt-5-6-sol-low",
          "verified_publish": true,
          "levels_completed": 1,
          "actions": 307,
          "resets": 3,
          "state": "GAME_OVER",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            31.517389966143426,
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            57,
            250,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            32,
            50,
            75,
            37,
            89,
            159,
            233,
            73
          ]
        },
        {
          "session_id": "22a25f67-1171-406a-99f7-74a0d00e76d8",
          "game_id": "ar25",
          "game_version": "0c556536",
          "published_at": "2026-05-29T14:16:09.363113Z",
          "score": 5.860061331094737,
          "model": "claude-opus-4-8",
          "runner": "benchmark_agent",
          "config": "anthropic-opus-4-8-high",
          "verified_publish": true,
          "levels_completed": 4,
          "actions": 1133,
          "resets": 21,
          "state": "GAME_OVER",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            115,
            5.260499957916001,
            4.364491274896999,
            18.08693354472189,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            24,
            218,
            359,
            87,
            445,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            32,
            50,
            75,
            37,
            89,
            159,
            233,
            73
          ]
        },
        {
          "session_id": "7752f433-a3a4-4026-ae8c-6c32b4afcf9d",
          "game_id": "ar25",
          "game_version": "0c556536",
          "published_at": "2026-05-01T16:12:33.823654Z",
          "score": 0.15378700499807768,
          "model": "claude-opus-4-7",
          "runner": "benchmark_agent",
          "config": "anthropic-opus-4-7-high",
          "verified_publish": true,
          "levels_completed": 1,
          "actions": 386,
          "resets": 9,
          "state": "GAME_OVER",
          "completed": false,
          "number_of_levels": 8,
          "level_scores": [
            5.536332179930796,
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            136,
            250,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            32,
            50,
            75,
            37,
            89,
            159,
            233,
            73
          ]
        },
        {
          "session_id": "c81599b9-d42a-4133-95af-8a108cd062f1",
          "game_id": "ar25",
          "game_version": "0c556536",
          "published_at": "2026-04-30T23:25:39.966909Z",
          "score": 0,
          "model": "gpt-5.5-2026-04-23",
          "runner": "benchmark_agent",
          "config": "openai-gpt-5-5-2026-04-23-high",
          "verified_publish": true,
          "levels_completed": 0,
          "actions": 160,
          "resets": 3,
          "state": "NOT_FINISHED",
          "completed": false,
          "number_of_levels": 8,
          "level_scores": [
            0,
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            160,
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            32,
            50,
            75,
            37,
            89,
            159,
            233,
            73
          ]
        },
        {
          "session_id": "a29335e6-9512-4f19-8f24-f53844553ca0",
          "game_id": "ar25",
          "game_version": "0c556536",
          "published_at": "2026-04-14T17:43:43.062565Z",
          "score": 0,
          "model": "grok-4.20-beta-0309-reasoning",
          "runner": "benchmark_agent",
          "config": "xai-grok-4-20-beta-0309-reasoning",
          "verified_publish": true,
          "levels_completed": 0,
          "actions": 101,
          "resets": 1,
          "state": "NOT_FINISHED",
          "completed": false,
          "number_of_levels": 8,
          "level_scores": [
            0,
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            101,
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            32,
            50,
            75,
            37,
            89,
            159,
            233,
            73
          ]
        },
        {
          "session_id": "b7efee02-3431-4704-bc0e-2ed7a9372e83",
          "game_id": "ar25",
          "game_version": "0c556536",
          "published_at": "2026-04-14T17:43:42.640115Z",
          "score": 0,
          "model": "gpt-5.4-2026-03-05",
          "runner": "benchmark_agent",
          "config": "openai-gpt-5-4-2026-03-05-high",
          "verified_publish": true,
          "levels_completed": 0,
          "actions": 160,
          "resets": 1,
          "state": "GAME_OVER",
          "completed": false,
          "number_of_levels": 8,
          "level_scores": [
            0,
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            160,
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            32,
            50,
            75,
            37,
            89,
            159,
            233,
            73
          ]
        },
        {
          "session_id": "52a24e10-2555-4609-a079-04cc129f4da3",
          "game_id": "ar25",
          "game_version": "0c556536",
          "published_at": "2026-04-14T17:43:41.825295Z",
          "score": 0,
          "model": "gemini-3.1-pro-preview",
          "runner": "benchmark_agent",
          "config": "google-gemini-3-1-pro-preview",
          "verified_publish": true,
          "levels_completed": 0,
          "actions": 110,
          "resets": 1,
          "state": "GAME_OVER",
          "completed": false,
          "number_of_levels": 8,
          "level_scores": [
            0,
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            110,
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            32,
            50,
            75,
            37,
            89,
            159,
            233,
            73
          ]
        },
        {
          "session_id": "bc829abb-1035-42b0-bdb9-18b24ba55a50",
          "game_id": "ar25",
          "game_version": "0c556536",
          "published_at": "2026-04-14T17:43:41.540666Z",
          "score": 0,
          "model": "claude-opus-4-6",
          "runner": "benchmark_agent",
          "config": "anthropic-opus-4-6-max-effort",
          "verified_publish": true,
          "levels_completed": 0,
          "actions": 160,
          "resets": 8,
          "state": "NOT_FINISHED",
          "completed": false,
          "number_of_levels": 8,
          "level_scores": [
            0,
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            160,
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            32,
            50,
            75,
            37,
            89,
            159,
            233,
            73
          ]
        }
      ]
    },
    {
      "game": "g50t",
      "source": "https://arcprize.org/tasks/g50t",
      "runs": [
        {
          "session_id": "45eb2e27-729e-4989-ba81-c2ceafa2841a",
          "game_id": "g50t",
          "game_version": "5849a774",
          "published_at": "2026-09-25T08:13:31.531546Z",
          "score": 3.571428571428571,
          "model": "gpt-6-sol",
          "runner": "benchmark_agent",
          "config": "openai-gpt-6-sol-none-provider-adapter",
          "verified_publish": true,
          "levels_completed": 1,
          "actions": 813,
          "resets": 23,
          "state": "GAME_OVER",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            115,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            45,
            768,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            78,
            175,
            179,
            230,
            96,
            54,
            67
          ]
        },
        {
          "session_id": "ac79f135-bf68-4243-9eb7-615a9cc002d1",
          "game_id": "g50t",
          "game_version": "5849a774",
          "published_at": "2026-09-25T07:48:26.467643Z",
          "score": 0,
          "model": "gpt-6-sol",
          "runner": "benchmark_agent",
          "config": "openai-gpt-6-sol-max",
          "verified_publish": true,
          "levels_completed": 0,
          "actions": 390,
          "resets": 3,
          "state": "GAME_OVER",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            390,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            78,
            175,
            179,
            230,
            96,
            54,
            67
          ]
        },
        {
          "session_id": "07d0720e-4b69-4e40-a449-adaccf2efa56",
          "game_id": "g50t",
          "game_version": "5849a774",
          "published_at": "2026-09-25T07:07:52.991744Z",
          "score": 0,
          "model": "gpt-6-sol",
          "runner": "benchmark_agent",
          "config": "openai-gpt-6-sol-low-provider-adapter",
          "verified_publish": true,
          "levels_completed": 0,
          "actions": 390,
          "resets": 16,
          "state": "NOT_FINISHED",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            390,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            78,
            175,
            179,
            230,
            96,
            54,
            67
          ]
        },
        {
          "session_id": "e5b509b9-0543-4a98-ab73-81de46e8fa82",
          "game_id": "g50t",
          "game_version": "5849a774",
          "published_at": "2026-09-25T06:44:50.224732Z",
          "score": 0,
          "model": "gpt-6-sol",
          "runner": "benchmark_agent",
          "config": "openai-gpt-6-sol-medium-provider-adapter",
          "verified_publish": true,
          "levels_completed": 0,
          "actions": 390,
          "resets": 14,
          "state": "GAME_OVER",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            390,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            78,
            175,
            179,
            230,
            96,
            54,
            67
          ]
        },
        {
          "session_id": "dc7913ea-1fa8-4d4c-89d9-431e43990942",
          "game_id": "g50t",
          "game_version": "5849a774",
          "published_at": "2026-09-25T06:40:17.186362Z",
          "score": 0,
          "model": "gpt-6-sol",
          "runner": "benchmark_agent",
          "config": "openai-gpt-6-sol-none",
          "verified_publish": true,
          "levels_completed": 0,
          "actions": 390,
          "resets": 2,
          "state": "GAME_OVER",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            390,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            78,
            175,
            179,
            230,
            96,
            54,
            67
          ]
        },
        {
          "session_id": "568addf3-6e5e-4a33-a849-d598550dbed1",
          "game_id": "g50t",
          "game_version": "5849a774",
          "published_at": "2026-09-25T06:15:16.809656Z",
          "score": 0,
          "model": "gpt-6-sol",
          "runner": "benchmark_agent",
          "config": "openai-gpt-6-sol-low",
          "verified_publish": true,
          "levels_completed": 0,
          "actions": 390,
          "resets": 4,
          "state": "GAME_OVER",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            390,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            78,
            175,
            179,
            230,
            96,
            54,
            67
          ]
        },
        {
          "session_id": "0799d8c9-530a-4196-b243-08631a0b9d95",
          "game_id": "g50t",
          "game_version": "5849a774",
          "published_at": "2026-09-25T06:05:19.175815Z",
          "score": 0,
          "model": "gpt-6-sol",
          "runner": "benchmark_agent",
          "config": "openai-gpt-6-sol-high-provider-adapter",
          "verified_publish": true,
          "levels_completed": 0,
          "actions": 390,
          "resets": 10,
          "state": "GAME_OVER",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            390,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            78,
            175,
            179,
            230,
            96,
            54,
            67
          ]
        },
        {
          "session_id": "5eb9ecfd-3b16-4bc5-836e-9c937d52e525",
          "game_id": "g50t",
          "game_version": "5849a774",
          "published_at": "2026-09-25T05:46:52.00962Z",
          "score": 0,
          "model": "gpt-6-sol",
          "runner": "benchmark_agent",
          "config": "openai-gpt-6-sol-xhigh-provider-adapter",
          "verified_publish": true,
          "levels_completed": 0,
          "actions": 390,
          "resets": 11,
          "state": "GAME_OVER",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            390,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            78,
            175,
            179,
            230,
            96,
            54,
            67
          ]
        },
        {
          "session_id": "be6cf7cd-9ac6-4582-a573-a9660cd9c501",
          "game_id": "g50t",
          "game_version": "5849a774",
          "published_at": "2026-09-25T05:43:55.61714Z",
          "score": 0,
          "model": "gpt-6-sol",
          "runner": "benchmark_agent",
          "config": "openai-gpt-6-sol-medium",
          "verified_publish": true,
          "levels_completed": 0,
          "actions": 390,
          "resets": 6,
          "state": "GAME_OVER",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            390,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            78,
            175,
            179,
            230,
            96,
            54,
            67
          ]
        },
        {
          "session_id": "705aba42-8f45-4869-88af-f76af09d87dc",
          "game_id": "g50t",
          "game_version": "5849a774",
          "published_at": "2026-09-25T05:24:39.246759Z",
          "score": 0,
          "model": "gpt-6-sol",
          "runner": "benchmark_agent",
          "config": "openai-gpt-6-sol-xhigh",
          "verified_publish": true,
          "levels_completed": 0,
          "actions": 390,
          "resets": 7,
          "state": "NOT_FINISHED",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            390,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            78,
            175,
            179,
            230,
            96,
            54,
            67
          ]
        },
        {
          "session_id": "7affc946-c62c-4739-8182-876fa60e9ec2",
          "game_id": "g50t",
          "game_version": "5849a774",
          "published_at": "2026-09-25T05:12:44.604422Z",
          "score": 0,
          "model": "gpt-6-sol",
          "runner": "benchmark_agent",
          "config": "openai-gpt-6-sol-high",
          "verified_publish": true,
          "levels_completed": 0,
          "actions": 390,
          "resets": 8,
          "state": "GAME_OVER",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            390,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            78,
            175,
            179,
            230,
            96,
            54,
            67
          ]
        },
        {
          "session_id": "2a370a2d-62c4-4bbc-9799-c95d4179d1ef",
          "game_id": "g50t",
          "game_version": "5849a774",
          "published_at": "2026-09-21T11:27:44.525599Z",
          "score": 1.8628747795414462,
          "model": "gemini-3.8-flash",
          "runner": "benchmark_agent",
          "config": "google-gemini-3-8-flash-low-provider-adapter",
          "verified_publish": true,
          "levels_completed": 1,
          "actions": 983,
          "resets": 24,
          "state": "GAME_OVER",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            52.160493827160494,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            108,
            875,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            78,
            175,
            179,
            230,
            96,
            54,
            67
          ]
        },
        {
          "session_id": "9738c85b-09e3-42df-a921-8a58c5cf4177",
          "game_id": "g50t",
          "game_version": "5849a774",
          "published_at": "2026-09-20T08:54:33.169249Z",
          "score": 10.714285714285714,
          "model": "gemini-3.8-flash",
          "runner": "benchmark_agent",
          "config": "google-gemini-3-8-flash-medium-provider-adapter",
          "verified_publish": true,
          "levels_completed": 2,
          "actions": 1104,
          "resets": 18,
          "state": "GAME_OVER",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            115,
            111.13732036580055,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            43,
            166,
            895,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            78,
            175,
            179,
            230,
            96,
            54,
            67
          ]
        },
        {
          "session_id": "323ceca8-0839-49a8-b26f-6a41b65bcd29",
          "game_id": "g50t",
          "game_version": "5849a774",
          "published_at": "2026-09-18T00:43:26.23957Z",
          "score": 62.691964285714285,
          "model": "gemini-3.8-flash",
          "runner": "benchmark_agent",
          "config": "google-gemini-3-8-flash-high-provider-adapter",
          "verified_publish": true,
          "levels_completed": 6,
          "actions": 929,
          "resets": 24,
          "state": "GAME_OVER",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            115,
            115,
            115,
            115,
            115,
            5.0625,
            0
          ],
          "level_actions": [
            30,
            47,
            161,
            31,
            85,
            240,
            335
          ],
          "level_baseline_actions": [
            78,
            175,
            179,
            230,
            96,
            54,
            67
          ]
        },
        {
          "session_id": "17aa0f6a-0e91-4f7b-be31-9210a0d9505a",
          "game_id": "g50t",
          "game_version": "5849a774",
          "published_at": "2026-09-16T20:37:04.091726Z",
          "score": 0,
          "model": "gpt-6-luna",
          "runner": "benchmark_agent",
          "config": "openai-gpt-6-luna-none-provider-adapter",
          "verified_publish": true,
          "levels_completed": 0,
          "actions": 390,
          "resets": 10,
          "state": "NOT_FINISHED",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            390,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            78,
            175,
            179,
            230,
            96,
            54,
            67
          ]
        },
        {
          "session_id": "7e41819c-5a85-4458-b255-faad442ba471",
          "game_id": "g50t",
          "game_version": "5849a774",
          "published_at": "2026-09-16T20:11:58.029269Z",
          "score": 0,
          "model": "gpt-6-luna",
          "runner": "benchmark_agent",
          "config": "openai-gpt-6-luna-medium",
          "verified_publish": true,
          "levels_completed": 0,
          "actions": 390,
          "resets": 2,
          "state": "GAME_OVER",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            390,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            78,
            175,
            179,
            230,
            96,
            54,
            67
          ]
        },
        {
          "session_id": "1a2b64c2-3a0d-4a53-bc17-d8fa7b1125d2",
          "game_id": "g50t",
          "game_version": "5849a774",
          "published_at": "2026-09-16T20:08:14.523183Z",
          "score": 0,
          "model": "gpt-6-luna",
          "runner": "benchmark_agent",
          "config": "openai-gpt-6-luna-medium-provider-adapter",
          "verified_publish": true,
          "levels_completed": 0,
          "actions": 390,
          "resets": 13,
          "state": "NOT_FINISHED",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            390,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            78,
            175,
            179,
            230,
            96,
            54,
            67
          ]
        },
        {
          "session_id": "f3ddb28a-fea2-4c60-8836-3ea4f1ec10bb",
          "game_id": "g50t",
          "game_version": "5849a774",
          "published_at": "2026-09-16T19:43:19.472635Z",
          "score": 0,
          "model": "gpt-6-luna",
          "runner": "benchmark_agent",
          "config": "openai-gpt-6-luna-low-provider-adapter",
          "verified_publish": true,
          "levels_completed": 0,
          "actions": 390,
          "resets": 7,
          "state": "GAME_OVER",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            390,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            78,
            175,
            179,
            230,
            96,
            54,
            67
          ]
        },
        {
          "session_id": "2f9dac95-5fd1-4908-aeed-ac327861998e",
          "game_id": "g50t",
          "game_version": "5849a774",
          "published_at": "2026-09-16T19:31:11.967543Z",
          "score": 0,
          "model": "gpt-6-luna",
          "runner": "benchmark_agent",
          "config": "openai-gpt-6-luna-none",
          "verified_publish": true,
          "levels_completed": 0,
          "actions": 390,
          "resets": 4,
          "state": "GAME_OVER",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            390,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            78,
            175,
            179,
            230,
            96,
            54,
            67
          ]
        },
        {
          "session_id": "c0a94f7d-f458-4ec4-a491-a46a97821b4c",
          "game_id": "g50t",
          "game_version": "5849a774",
          "published_at": "2026-09-16T17:56:59.331606Z",
          "score": 0,
          "model": "gpt-6-luna",
          "runner": "benchmark_agent",
          "config": "openai-gpt-6-luna-low",
          "verified_publish": true,
          "levels_completed": 0,
          "actions": 390,
          "resets": 2,
          "state": "GAME_OVER",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            390,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            78,
            175,
            179,
            230,
            96,
            54,
            67
          ]
        },
        {
          "session_id": "08e2b89f-b368-40b6-931c-cbb1f2a004a1",
          "game_id": "g50t",
          "game_version": "5849a774",
          "published_at": "2026-09-16T17:04:41.974578Z",
          "score": 0,
          "model": "gpt-6-luna",
          "runner": "benchmark_agent",
          "config": "openai-gpt-6-luna-high",
          "verified_publish": true,
          "levels_completed": 0,
          "actions": 390,
          "resets": 2,
          "state": "GAME_OVER",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            390,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            78,
            175,
            179,
            230,
            96,
            54,
            67
          ]
        },
        {
          "session_id": "a300de13-2c70-4bb5-993f-c68d07e8bb23",
          "game_id": "g50t",
          "game_version": "5849a774",
          "published_at": "2026-09-16T16:38:42.07168Z",
          "score": 0,
          "model": "gpt-6-luna",
          "runner": "benchmark_agent",
          "config": "openai-gpt-6-luna-high-provider-adapter",
          "verified_publish": true,
          "levels_completed": 0,
          "actions": 390,
          "resets": 11,
          "state": "GAME_OVER",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            390,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            78,
            175,
            179,
            230,
            96,
            54,
            67
          ]
        },
        {
          "session_id": "26ee30e7-44e5-496c-b7e2-f15bbce4c447",
          "game_id": "g50t",
          "game_version": "5849a774",
          "published_at": "2026-09-16T12:10:58.261625Z",
          "score": 0,
          "model": "gpt-6-luna",
          "runner": "benchmark_agent",
          "config": "openai-gpt-6-luna-max",
          "verified_publish": true,
          "levels_completed": 0,
          "actions": 390,
          "resets": 3,
          "state": "GAME_OVER",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            390,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            78,
            175,
            179,
            230,
            96,
            54,
            67
          ]
        },
        {
          "session_id": "e9673e5a-eda7-46e7-810b-959397a8c851",
          "game_id": "g50t",
          "game_version": "5849a774",
          "published_at": "2026-09-16T09:19:38.675762Z",
          "score": 0,
          "model": "gpt-6-luna",
          "runner": "benchmark_agent",
          "config": "openai-gpt-6-luna-xhigh",
          "verified_publish": true,
          "levels_completed": 0,
          "actions": 390,
          "resets": 3,
          "state": "GAME_OVER",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            390,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            78,
            175,
            179,
            230,
            96,
            54,
            67
          ]
        },
        {
          "session_id": "8d6cf3f8-549c-4cbc-a816-6e2cf127b6d8",
          "game_id": "g50t",
          "game_version": "5849a774",
          "published_at": "2026-09-16T09:09:48.576077Z",
          "score": 0.28115226215738603,
          "model": "gpt-6-1-sol",
          "runner": "benchmark_agent",
          "config": "openai-gpt-6-1-sol-xhigh",
          "verified_publish": true,
          "levels_completed": 1,
          "actions": 1084,
          "resets": 8,
          "state": "GAME_OVER",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            7.87226334040681,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            278,
            806,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            78,
            175,
            179,
            230,
            96,
            54,
            67
          ]
        },
        {
          "session_id": "7d5bc55d-9e1b-474d-88b0-95394a55f5aa",
          "game_id": "g50t",
          "game_version": "5849a774",
          "published_at": "2026-09-16T08:54:34.683283Z",
          "score": 0,
          "model": "gpt-6-luna",
          "runner": "benchmark_agent",
          "config": "openai-gpt-6-luna-xhigh-provider-adapter",
          "verified_publish": true,
          "levels_completed": 0,
          "actions": 390,
          "resets": 18,
          "state": "GAME_OVER",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            390,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            78,
            175,
            179,
            230,
            96,
            54,
            67
          ]
        },
        {
          "session_id": "c322dbc5-0164-41fe-bfce-95af3b741a6e",
          "game_id": "g50t",
          "game_version": "5849a774",
          "published_at": "2026-09-16T07:11:04.979534Z",
          "score": 0,
          "model": "gpt-6-1-sol",
          "runner": "benchmark_agent",
          "config": "openai-gpt-6-1-sol-low",
          "verified_publish": true,
          "levels_completed": 0,
          "actions": 390,
          "resets": 2,
          "state": "GAME_OVER",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            390,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            78,
            175,
            179,
            230,
            96,
            54,
            67
          ]
        },
        {
          "session_id": "968592dd-dcab-4b46-9324-34e4097b8309",
          "game_id": "g50t",
          "game_version": "5849a774",
          "published_at": "2026-09-16T06:50:09.991771Z",
          "score": 0,
          "model": "gpt-6-1-sol",
          "runner": "benchmark_agent",
          "config": "openai-gpt-6-1-sol-max",
          "verified_publish": true,
          "levels_completed": 0,
          "actions": 390,
          "resets": 3,
          "state": "GAME_OVER",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            390,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            78,
            175,
            179,
            230,
            96,
            54,
            67
          ]
        },
        {
          "session_id": "83184a70-938c-40a1-b1c3-744cf823ac50",
          "game_id": "g50t",
          "game_version": "5849a774",
          "published_at": "2026-09-16T05:52:58.141787Z",
          "score": 100,
          "model": "gpt-6-1-sol",
          "runner": "benchmark_agent",
          "config": "openai-gpt-6-1-sol-high-provider-adapter",
          "verified_publish": true,
          "levels_completed": 7,
          "actions": 306,
          "resets": 1,
          "state": "WIN",
          "completed": true,
          "number_of_levels": 0,
          "level_scores": [
            115,
            115,
            115,
            115,
            115,
            115,
            115
          ],
          "level_actions": [
            24,
            31,
            71,
            31,
            57,
            49,
            43
          ],
          "level_baseline_actions": [
            78,
            175,
            179,
            230,
            96,
            54,
            67
          ]
        },
        {
          "session_id": "e1cf861c-ca81-4b0f-8d6c-3b487d6a745e",
          "game_id": "g50t",
          "game_version": "5849a774",
          "published_at": "2026-09-16T05:52:56.872993Z",
          "score": 0,
          "model": "gpt-6-1-sol",
          "runner": "benchmark_agent",
          "config": "openai-gpt-6-1-sol-medium",
          "verified_publish": true,
          "levels_completed": 0,
          "actions": 390,
          "resets": 3,
          "state": "GAME_OVER",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            390,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            78,
            175,
            179,
            230,
            96,
            54,
            67
          ]
        },
        {
          "session_id": "db4c1ea2-687c-4013-b50d-1a7a0710308c",
          "game_id": "g50t",
          "game_version": "5849a774",
          "published_at": "2026-09-16T05:35:43.708966Z",
          "score": 100,
          "model": "gpt-6-1-sol",
          "runner": "benchmark_agent",
          "config": "openai-gpt-6-1-sol-max-provider-adapter",
          "verified_publish": true,
          "levels_completed": 7,
          "actions": 304,
          "resets": 1,
          "state": "WIN",
          "completed": true,
          "number_of_levels": 0,
          "level_scores": [
            115,
            115,
            115,
            115,
            115,
            115,
            115
          ],
          "level_actions": [
            27,
            31,
            72,
            31,
            50,
            50,
            43
          ],
          "level_baseline_actions": [
            78,
            175,
            179,
            230,
            96,
            54,
            67
          ]
        },
        {
          "session_id": "8c2010ca-7aec-499e-b556-8f16306f1d8b",
          "game_id": "g50t",
          "game_version": "5849a774",
          "published_at": "2026-09-16T05:19:11.091582Z",
          "score": 100,
          "model": "gpt-6-1-sol",
          "runner": "benchmark_agent",
          "config": "openai-gpt-6-1-sol-xhigh-provider-adapter",
          "verified_publish": true,
          "levels_completed": 7,
          "actions": 379,
          "resets": 0,
          "state": "WIN",
          "completed": true,
          "number_of_levels": 0,
          "level_scores": [
            115,
            115,
            115,
            115,
            106.55567117585846,
            115,
            115
          ],
          "level_actions": [
            21,
            31,
            74,
            68,
            93,
            49,
            43
          ],
          "level_baseline_actions": [
            78,
            175,
            179,
            230,
            96,
            54,
            67
          ]
        },
        {
          "session_id": "91bce332-642d-4eaa-8bbd-8c37c3540bc4",
          "game_id": "g50t",
          "game_version": "5849a774",
          "published_at": "2026-09-16T04:36:48.794245Z",
          "score": 0,
          "model": "gpt-6-luna",
          "runner": "benchmark_agent",
          "config": "openai-gpt-6-luna-max-provider-adapter",
          "verified_publish": true,
          "levels_completed": 0,
          "actions": 390,
          "resets": 23,
          "state": "GAME_OVER",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            390,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            78,
            175,
            179,
            230,
            96,
            54,
            67
          ]
        },
        {
          "session_id": "efb2848f-9ae9-4a5d-99bf-323f679ee862",
          "game_id": "g50t",
          "game_version": "5849a774",
          "published_at": "2026-09-16T04:28:12.980972Z",
          "score": 0,
          "model": "gpt-6-1-sol",
          "runner": "benchmark_agent",
          "config": "openai-gpt-6-1-sol-high",
          "verified_publish": true,
          "levels_completed": 0,
          "actions": 390,
          "resets": 3,
          "state": "GAME_OVER",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            390,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            78,
            175,
            179,
            230,
            96,
            54,
            67
          ]
        },
        {
          "session_id": "cf0fed3e-812e-4b69-8b43-b9dac8484289",
          "game_id": "g50t",
          "game_version": "5849a774",
          "published_at": "2026-09-12T05:11:23.343757Z",
          "score": 0,
          "model": "gemini-3.8-flash",
          "runner": "benchmark_agent",
          "config": "google-gemini-3-8-flash-low",
          "verified_publish": true,
          "levels_completed": 0,
          "actions": 390,
          "resets": 3,
          "state": "GAME_OVER",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            390,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            78,
            175,
            179,
            230,
            96,
            54,
            67
          ]
        },
        {
          "session_id": "0d230966-f913-4751-9154-7ae73ad4102b",
          "game_id": "g50t",
          "game_version": "5849a774",
          "published_at": "2026-09-11T01:59:40.213579Z",
          "score": 0,
          "model": "gemini-3.8-flash",
          "runner": "benchmark_agent",
          "config": "google-gemini-3-8-flash-high",
          "verified_publish": true,
          "levels_completed": 0,
          "actions": 390,
          "resets": 2,
          "state": "GAME_OVER",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            390,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            78,
            175,
            179,
            230,
            96,
            54,
            67
          ]
        },
        {
          "session_id": "49a93459-4457-414b-ac40-2a478d5786d7",
          "game_id": "g50t",
          "game_version": "5849a774",
          "published_at": "2026-09-10T20:56:32.235399Z",
          "score": 0,
          "model": "gemini-3.8-flash",
          "runner": "benchmark_agent",
          "config": "google-gemini-3-8-flash-medium",
          "verified_publish": true,
          "levels_completed": 0,
          "actions": 390,
          "resets": 2,
          "state": "GAME_OVER",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            390,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            78,
            175,
            179,
            230,
            96,
            54,
            67
          ]
        },
        {
          "session_id": "126c9eeb-485a-4982-9a99-f18d4ebb4f55",
          "game_id": "g50t",
          "game_version": "5849a774",
          "published_at": "2026-08-11T12:08:26.305212Z",
          "score": 0,
          "model": "grok-4.6",
          "runner": "benchmark_agent",
          "config": "xai-grok-4-6-xhigh",
          "verified_publish": true,
          "levels_completed": 0,
          "actions": 390,
          "resets": 8,
          "state": "NOT_FINISHED",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            390,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            78,
            175,
            179,
            230,
            96,
            54,
            67
          ]
        },
        {
          "session_id": "9ec9ae2e-9625-40ce-bfa1-89ac0756c683",
          "game_id": "g50t",
          "game_version": "5849a774",
          "published_at": "2026-07-21T09:17:08.670387Z",
          "score": 0,
          "model": "claude-opus-5",
          "runner": "benchmark_agent",
          "config": "anthropic-claude-opus-5-high",
          "verified_publish": true,
          "levels_completed": 0,
          "actions": 390,
          "resets": 2,
          "state": "GAME_OVER",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            390,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            78,
            175,
            179,
            230,
            96,
            54,
            67
          ]
        },
        {
          "session_id": "781a9428-86ac-4f7d-9df8-e77dfd57b8e6",
          "game_id": "g50t",
          "game_version": "5849a774",
          "published_at": "2026-06-24T11:04:12.242931Z",
          "score": 0,
          "model": "gpt-5-6-luna",
          "runner": "benchmark_agent",
          "config": "openai-gpt-5-6-luna-low",
          "verified_publish": true,
          "levels_completed": 0,
          "actions": 390,
          "resets": 2,
          "state": "GAME_OVER",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            390,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            78,
            175,
            179,
            230,
            96,
            54,
            67
          ]
        },
        {
          "session_id": "bb4ced4d-9329-47a3-8c82-1a4856b34e42",
          "game_id": "g50t",
          "game_version": "5849a774",
          "published_at": "2026-06-24T11:03:39.767874Z",
          "score": 0,
          "model": "gpt-5-6-luna",
          "runner": "benchmark_agent",
          "config": "openai-gpt-5-6-luna-medium",
          "verified_publish": true,
          "levels_completed": 0,
          "actions": 390,
          "resets": 2,
          "state": "GAME_OVER",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            390,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            78,
            175,
            179,
            230,
            96,
            54,
            67
          ]
        },
        {
          "session_id": "6d4c964f-6bfa-4d96-b0f2-985ee5c41a18",
          "game_id": "g50t",
          "game_version": "5849a774",
          "published_at": "2026-06-24T06:14:38.884551Z",
          "score": 0,
          "model": "gpt-5-6-luna",
          "runner": "benchmark_agent",
          "config": "openai-gpt-5-6-luna-xhigh",
          "verified_publish": true,
          "levels_completed": 0,
          "actions": 390,
          "resets": 14,
          "state": "NOT_FINISHED",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            390,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            78,
            175,
            179,
            230,
            96,
            54,
            67
          ]
        },
        {
          "session_id": "0091794f-99eb-4be4-b890-d55a22334def",
          "game_id": "g50t",
          "game_version": "5849a774",
          "published_at": "2026-06-24T05:52:28.10783Z",
          "score": 0,
          "model": "gpt-5-6-luna",
          "runner": "benchmark_agent",
          "config": "openai-gpt-5-6-luna-high",
          "verified_publish": true,
          "levels_completed": 0,
          "actions": 390,
          "resets": 5,
          "state": "GAME_OVER",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            390,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            78,
            175,
            179,
            230,
            96,
            54,
            67
          ]
        },
        {
          "session_id": "c4a0dc33-0436-4580-85ed-9ad60062e93e",
          "game_id": "g50t",
          "game_version": "5849a774",
          "published_at": "2026-06-23T06:49:21.227089Z",
          "score": 0,
          "model": "gpt-5-6-luna",
          "runner": "benchmark_agent",
          "config": "openai-gpt-5-6-luna-max",
          "verified_publish": true,
          "levels_completed": 0,
          "actions": 390,
          "resets": 59,
          "state": "NOT_FINISHED",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            390,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            78,
            175,
            179,
            230,
            96,
            54,
            67
          ]
        },
        {
          "session_id": "9542d7e4-3f94-4d1c-94a5-6328a0fdeab5",
          "game_id": "g50t",
          "game_version": "5849a774",
          "published_at": "2026-06-20T12:12:41.190766Z",
          "score": 0,
          "model": "gpt-5-6-terra",
          "runner": "benchmark_agent",
          "config": "openai-gpt-5-6-terra-low",
          "verified_publish": true,
          "levels_completed": 0,
          "actions": 390,
          "resets": 5,
          "state": "GAME_OVER",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            390,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            78,
            175,
            179,
            230,
            96,
            54,
            67
          ]
        },
        {
          "session_id": "d2cc0e3f-dc13-45c5-8bd4-cc3e25013cdd",
          "game_id": "g50t",
          "game_version": "5849a774",
          "published_at": "2026-06-20T09:42:27.596677Z",
          "score": 0,
          "model": "gpt-5-6-terra",
          "runner": "benchmark_agent",
          "config": "openai-gpt-5-6-terra-medium",
          "verified_publish": true,
          "levels_completed": 0,
          "actions": 390,
          "resets": 7,
          "state": "GAME_OVER",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            390,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            78,
            175,
            179,
            230,
            96,
            54,
            67
          ]
        },
        {
          "session_id": "1c0a190c-2c78-4293-8af4-448e29fc57e2",
          "game_id": "g50t",
          "game_version": "5849a774",
          "published_at": "2026-06-20T09:23:36.397079Z",
          "score": 0,
          "model": "gpt-5-6-terra",
          "runner": "benchmark_agent",
          "config": "openai-gpt-5-6-terra-high",
          "verified_publish": true,
          "levels_completed": 0,
          "actions": 390,
          "resets": 11,
          "state": "NOT_FINISHED",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            390,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            78,
            175,
            179,
            230,
            96,
            54,
            67
          ]
        },
        {
          "session_id": "acdf2bad-3e41-4885-8798-d1a65fecd540",
          "game_id": "g50t",
          "game_version": "5849a774",
          "published_at": "2026-06-20T04:35:30.514519Z",
          "score": 0,
          "model": "gpt-5-6-terra",
          "runner": "benchmark_agent",
          "config": "openai-gpt-5-6-terra-xhigh",
          "verified_publish": true,
          "levels_completed": 0,
          "actions": 390,
          "resets": 8,
          "state": "NOT_FINISHED",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            390,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            78,
            175,
            179,
            230,
            96,
            54,
            67
          ]
        },
        {
          "session_id": "8f2ef952-300b-4122-a249-e3e45007590f",
          "game_id": "g50t",
          "game_version": "5849a774",
          "published_at": "2026-06-19T02:33:01.361743Z",
          "score": 0,
          "model": "gpt-5-6-terra",
          "runner": "benchmark_agent",
          "config": "openai-gpt-5-6-terra-max",
          "verified_publish": true,
          "levels_completed": 0,
          "actions": 390,
          "resets": 9,
          "state": "NOT_FINISHED",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            390,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            78,
            175,
            179,
            230,
            96,
            54,
            67
          ]
        },
        {
          "session_id": "d7c19632-ebc5-4162-8d19-5e1a1a3d000d",
          "game_id": "g50t",
          "game_version": "5849a774",
          "published_at": "2026-06-13T09:45:36.031806Z",
          "score": 0,
          "model": "gpt-5-6-sol",
          "runner": "benchmark_agent",
          "config": "openai-gpt-5-6-sol-max",
          "verified_publish": true,
          "levels_completed": 0,
          "actions": 390,
          "resets": 4,
          "state": "NOT_FINISHED",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            390,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            78,
            175,
            179,
            230,
            96,
            54,
            67
          ]
        },
        {
          "session_id": "a590485b-2dfa-407b-8205-63068d57fb5e",
          "game_id": "g50t",
          "game_version": "5849a774",
          "published_at": "2026-06-12T04:27:15.90076Z",
          "score": 0,
          "model": "gpt-5-6-sol",
          "runner": "benchmark_agent",
          "config": "openai-gpt-5-6-sol-medium",
          "verified_publish": true,
          "levels_completed": 0,
          "actions": 390,
          "resets": 4,
          "state": "GAME_OVER",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            390,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            78,
            175,
            179,
            230,
            96,
            54,
            67
          ]
        },
        {
          "session_id": "9fa0cae5-52e3-4f4e-a4ce-cff8f4aadabb",
          "game_id": "g50t",
          "game_version": "5849a774",
          "published_at": "2026-06-11T21:19:13.190688Z",
          "score": 0,
          "model": "gpt-5-6-sol",
          "runner": "benchmark_agent",
          "config": "openai-gpt-5-6-sol-xhigh",
          "verified_publish": true,
          "levels_completed": 0,
          "actions": 390,
          "resets": 6,
          "state": "GAME_OVER",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            390,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            78,
            175,
            179,
            230,
            96,
            54,
            67
          ]
        },
        {
          "session_id": "c96c1aa1-88cb-4ed6-a313-2debad2740f3",
          "game_id": "g50t",
          "game_version": "5849a774",
          "published_at": "2026-06-10T14:38:19.096089Z",
          "score": 0,
          "model": "gpt-5-6-sol",
          "runner": "benchmark_agent",
          "config": "openai-gpt-5-6-sol-high",
          "verified_publish": true,
          "levels_completed": 0,
          "actions": 390,
          "resets": 4,
          "state": "GAME_OVER",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            390,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            78,
            175,
            179,
            230,
            96,
            54,
            67
          ]
        },
        {
          "session_id": "a546a03b-94ab-419e-aad6-c69736161f71",
          "game_id": "g50t",
          "game_version": "5849a774",
          "published_at": "2026-06-09T22:27:38.288393Z",
          "score": 0,
          "model": "gpt-5-6-sol",
          "runner": "benchmark_agent",
          "config": "openai-gpt-5-6-sol-low",
          "verified_publish": true,
          "levels_completed": 0,
          "actions": 390,
          "resets": 4,
          "state": "GAME_OVER",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            390,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            78,
            175,
            179,
            230,
            96,
            54,
            67
          ]
        },
        {
          "session_id": "7788c1e6-c567-484b-9a66-3fee8972fbb3",
          "game_id": "g50t",
          "game_version": "5849a774",
          "published_at": "2026-05-30T23:22:09.305338Z",
          "score": 0,
          "model": "claude-opus-4-8",
          "runner": "benchmark_agent",
          "config": "anthropic-opus-4-8-high",
          "verified_publish": true,
          "levels_completed": 0,
          "actions": 390,
          "resets": 3,
          "state": "GAME_OVER",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            390,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            78,
            175,
            179,
            230,
            96,
            54,
            67
          ]
        },
        {
          "session_id": "70ecc44d-2fda-4c15-8e3f-c9d29e067943",
          "game_id": "g50t",
          "game_version": "5849a774",
          "published_at": "2026-05-01T05:09:05.663393Z",
          "score": 0,
          "model": "claude-opus-4-7",
          "runner": "benchmark_agent",
          "config": "anthropic-opus-4-7-high",
          "verified_publish": true,
          "levels_completed": 0,
          "actions": 390,
          "resets": 2,
          "state": "GAME_OVER",
          "completed": false,
          "number_of_levels": 7,
          "level_scores": [
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            390,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            78,
            175,
            179,
            230,
            96,
            54,
            67
          ]
        },
        {
          "session_id": "d90b94c8-52b0-4f6e-b0e6-0dc02bf21e06",
          "game_id": "g50t",
          "game_version": "5849a774",
          "published_at": "2026-04-30T23:23:07.316461Z",
          "score": 0,
          "model": "gpt-5.5-2026-04-23",
          "runner": "benchmark_agent",
          "config": "openai-gpt-5-5-2026-04-23-high",
          "verified_publish": true,
          "levels_completed": 0,
          "actions": 390,
          "resets": 4,
          "state": "GAME_OVER",
          "completed": false,
          "number_of_levels": 7,
          "level_scores": [
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            390,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            78,
            175,
            179,
            230,
            96,
            54,
            67
          ]
        },
        {
          "session_id": "a26ed3f0-5a3b-42bc-929c-6e5f7bbc75bc",
          "game_id": "g50t",
          "game_version": "5849a774",
          "published_at": "2026-04-14T17:44:26.391615Z",
          "score": 0,
          "model": "grok-4.20-beta-0309-reasoning",
          "runner": "benchmark_agent",
          "config": "xai-grok-4-20-beta-0309-reasoning",
          "verified_publish": true,
          "levels_completed": 0,
          "actions": 129,
          "resets": 6,
          "state": "NOT_FINISHED",
          "completed": false,
          "number_of_levels": 7,
          "level_scores": [
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            129,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            78,
            175,
            179,
            230,
            96,
            54,
            67
          ]
        },
        {
          "session_id": "6c9c2310-e96a-4cb7-adbc-0e231407c594",
          "game_id": "g50t",
          "game_version": "5849a774",
          "published_at": "2026-04-14T17:44:25.201071Z",
          "score": 0,
          "model": "gpt-5.4-2026-03-05",
          "runner": "benchmark_agent",
          "config": "openai-gpt-5-4-2026-03-05-high",
          "verified_publish": true,
          "levels_completed": 0,
          "actions": 390,
          "resets": 2,
          "state": "GAME_OVER",
          "completed": false,
          "number_of_levels": 7,
          "level_scores": [
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            390,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            78,
            175,
            179,
            230,
            96,
            54,
            67
          ]
        },
        {
          "session_id": "49e4da0c-be5c-4f55-b75b-8c952a0c5a0f",
          "game_id": "g50t",
          "game_version": "5849a774",
          "published_at": "2026-04-14T17:44:24.306063Z",
          "score": 0,
          "model": "gemini-3.1-pro-preview",
          "runner": "benchmark_agent",
          "config": "google-gemini-3-1-pro-preview",
          "verified_publish": true,
          "levels_completed": 0,
          "actions": 390,
          "resets": 4,
          "state": "GAME_OVER",
          "completed": false,
          "number_of_levels": 7,
          "level_scores": [
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            390,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            78,
            175,
            179,
            230,
            96,
            54,
            67
          ]
        },
        {
          "session_id": "ecf94cb0-8c5f-4508-a223-9ad92acf0cf6",
          "game_id": "g50t",
          "game_version": "5849a774",
          "published_at": "2026-04-14T17:44:23.298452Z",
          "score": 0,
          "model": "claude-opus-4-6",
          "runner": "benchmark_agent",
          "config": "anthropic-opus-4-6-max-effort",
          "verified_publish": true,
          "levels_completed": 0,
          "actions": 131,
          "resets": 1,
          "state": "GAME_OVER",
          "completed": false,
          "number_of_levels": 7,
          "level_scores": [
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            131,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            78,
            175,
            179,
            230,
            96,
            54,
            67
          ]
        }
      ]
    },
    {
      "game": "ls20",
      "source": "https://arcprize.org/tasks/ls20",
      "runs": [
        {
          "session_id": "ad51250c-7e97-4fff-b3ce-d728bca549cb",
          "game_id": "ls20",
          "game_version": "9607627b",
          "published_at": "2026-09-25T15:27:30.10973Z",
          "score": 0.8169052119902782,
          "model": "gpt-6-sol",
          "runner": "benchmark_agent",
          "config": "openai-gpt-6-sol-max",
          "verified_publish": true,
          "levels_completed": 1,
          "actions": 589,
          "resets": 6,
          "state": "GAME_OVER",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            22.87334593572779,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            46,
            543,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            22,
            123,
            73,
            84,
            96,
            192,
            186
          ]
        },
        {
          "session_id": "40c677c4-cf6e-4aca-b9e1-c7f168010ecd",
          "game_id": "ls20",
          "game_version": "9607627b",
          "published_at": "2026-09-25T07:06:04.016234Z",
          "score": 0.49657323429228056,
          "model": "gpt-6-sol",
          "runner": "benchmark_agent",
          "config": "openai-gpt-6-sol-high",
          "verified_publish": true,
          "levels_completed": 1,
          "actions": 674,
          "resets": 6,
          "state": "GAME_OVER",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            13.904050560183856,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            59,
            615,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            22,
            123,
            73,
            84,
            96,
            192,
            186
          ]
        },
        {
          "session_id": "d01065cc-968a-425c-b4fb-f3cda4482c74",
          "game_id": "ls20",
          "game_version": "9607627b",
          "published_at": "2026-09-25T06:53:29.212556Z",
          "score": 0.4355181225929525,
          "model": "gpt-6-sol",
          "runner": "benchmark_agent",
          "config": "openai-gpt-6-sol-low-provider-adapter",
          "verified_publish": true,
          "levels_completed": 1,
          "actions": 678,
          "resets": 7,
          "state": "GAME_OVER",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            12.19450743260267,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            63,
            615,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            22,
            123,
            73,
            84,
            96,
            192,
            186
          ]
        },
        {
          "session_id": "23b50e2c-9509-4937-90b6-f2df9033ff65",
          "game_id": "ls20",
          "game_version": "9607627b",
          "published_at": "2026-09-25T06:35:55.02056Z",
          "score": 0,
          "model": "gpt-6-sol",
          "runner": "benchmark_agent",
          "config": "openai-gpt-6-sol-none-provider-adapter",
          "verified_publish": true,
          "levels_completed": 0,
          "actions": 110,
          "resets": 0,
          "state": "NOT_FINISHED",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            110,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            22,
            123,
            73,
            84,
            96,
            192,
            186
          ]
        },
        {
          "session_id": "749e6624-3d81-4ac1-a921-18732967e050",
          "game_id": "ls20",
          "game_version": "9607627b",
          "published_at": "2026-09-25T06:32:36.905463Z",
          "score": 3.571428571428571,
          "model": "gpt-6-sol",
          "runner": "benchmark_agent",
          "config": "openai-gpt-6-sol-medium-provider-adapter",
          "verified_publish": true,
          "levels_completed": 1,
          "actions": 635,
          "resets": 4,
          "state": "GAME_OVER",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            115,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            20,
            615,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            22,
            123,
            73,
            84,
            96,
            192,
            186
          ]
        },
        {
          "session_id": "eed1e7fd-af4e-49b7-a763-ff3a205377f1",
          "game_id": "ls20",
          "game_version": "9607627b",
          "published_at": "2026-09-25T06:25:04.703146Z",
          "score": 0,
          "model": "gpt-6-sol",
          "runner": "benchmark_agent",
          "config": "openai-gpt-6-sol-none",
          "verified_publish": true,
          "levels_completed": 0,
          "actions": 110,
          "resets": 0,
          "state": "NOT_FINISHED",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            110,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            22,
            123,
            73,
            84,
            96,
            192,
            186
          ]
        },
        {
          "session_id": "b7fdc0f8-bdf2-457a-adb6-b93c2ebdd50a",
          "game_id": "ls20",
          "game_version": "9607627b",
          "published_at": "2026-09-25T05:59:37.473878Z",
          "score": 0,
          "model": "gpt-6-sol",
          "runner": "benchmark_agent",
          "config": "openai-gpt-6-sol-low",
          "verified_publish": true,
          "levels_completed": 0,
          "actions": 110,
          "resets": 5,
          "state": "NOT_FINISHED",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            110,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            22,
            123,
            73,
            84,
            96,
            192,
            186
          ]
        },
        {
          "session_id": "a8e5dd8b-4ee7-4922-a3f4-652f86fd0483",
          "game_id": "ls20",
          "game_version": "9607627b",
          "published_at": "2026-09-25T05:45:14.746634Z",
          "score": 3.571428571428571,
          "model": "gpt-6-sol",
          "runner": "benchmark_agent",
          "config": "openai-gpt-6-sol-high-provider-adapter",
          "verified_publish": true,
          "levels_completed": 1,
          "actions": 636,
          "resets": 6,
          "state": "GAME_OVER",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            109.75056689342404,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            21,
            615,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            22,
            123,
            73,
            84,
            96,
            192,
            186
          ]
        },
        {
          "session_id": "c2026142-69d3-4a33-ac92-c324b5035720",
          "game_id": "ls20",
          "game_version": "9607627b",
          "published_at": "2026-09-25T05:25:57.690704Z",
          "score": 0,
          "model": "gpt-6-sol",
          "runner": "benchmark_agent",
          "config": "openai-gpt-6-sol-medium",
          "verified_publish": true,
          "levels_completed": 0,
          "actions": 110,
          "resets": 0,
          "state": "NOT_FINISHED",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            110,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            22,
            123,
            73,
            84,
            96,
            192,
            186
          ]
        },
        {
          "session_id": "5e766fe9-e9dd-4e38-b520-881a549cc7bd",
          "game_id": "ls20",
          "game_version": "9607627b",
          "published_at": "2026-09-25T05:10:36.025916Z",
          "score": 2.5570583262890954,
          "model": "gpt-6-sol",
          "runner": "benchmark_agent",
          "config": "openai-gpt-6-sol-xhigh-provider-adapter",
          "verified_publish": true,
          "levels_completed": 1,
          "actions": 641,
          "resets": 4,
          "state": "GAME_OVER",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            71.59763313609467,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            26,
            615,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            22,
            123,
            73,
            84,
            96,
            192,
            186
          ]
        },
        {
          "session_id": "9ec6cf12-5a0b-4468-9192-fa0538456997",
          "game_id": "ls20",
          "game_version": "9607627b",
          "published_at": "2026-09-25T04:28:58.201568Z",
          "score": 0,
          "model": "gpt-6-sol",
          "runner": "benchmark_agent",
          "config": "openai-gpt-6-sol-xhigh",
          "verified_publish": true,
          "levels_completed": 0,
          "actions": 110,
          "resets": 2,
          "state": "NOT_FINISHED",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            110,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            22,
            123,
            73,
            84,
            96,
            192,
            186
          ]
        },
        {
          "session_id": "f85c1d39-7f86-491b-b353-92da72fea2e4",
          "game_id": "ls20",
          "game_version": "9607627b",
          "published_at": "2026-09-21T11:28:38.150756Z",
          "score": 2.5570583262890954,
          "model": "gemini-3.8-flash",
          "runner": "benchmark_agent",
          "config": "google-gemini-3-8-flash-low-provider-adapter",
          "verified_publish": true,
          "levels_completed": 1,
          "actions": 641,
          "resets": 6,
          "state": "GAME_OVER",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            71.59763313609467,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            26,
            615,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            22,
            123,
            73,
            84,
            96,
            192,
            186
          ]
        },
        {
          "session_id": "c5abe636-7576-4f84-b979-d003d81a9d2f",
          "game_id": "ls20",
          "game_version": "9607627b",
          "published_at": "2026-09-20T05:57:49.503059Z",
          "score": 4.689259115223628,
          "model": "gemini-3.8-flash",
          "runner": "benchmark_agent",
          "config": "google-gemini-3-8-flash-medium-provider-adapter",
          "verified_publish": true,
          "levels_completed": 2,
          "actions": 642,
          "resets": 5,
          "state": "GAME_OVER",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            84.02777777777777,
            23.635738724241907,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            24,
            253,
            365,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            22,
            123,
            73,
            84,
            96,
            192,
            186
          ]
        },
        {
          "session_id": "b75e1e99-e10c-42b0-874c-e07dbea51165",
          "game_id": "ls20",
          "game_version": "9607627b",
          "published_at": "2026-09-17T17:25:48.460325Z",
          "score": 5.416662601110384,
          "model": "gemini-3.8-flash",
          "runner": "benchmark_agent",
          "config": "google-gemini-3-8-flash-high-provider-adapter",
          "verified_publish": true,
          "levels_completed": 2,
          "actions": 629,
          "resets": 4,
          "state": "GAME_OVER",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            100,
            25.83327641554538,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            22,
            242,
            365,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            22,
            123,
            73,
            84,
            96,
            192,
            186
          ]
        },
        {
          "session_id": "00da04df-6efd-421e-837b-e2b1221487c4",
          "game_id": "ls20",
          "game_version": "9607627b",
          "published_at": "2026-09-16T19:24:43.696717Z",
          "score": 0,
          "model": "gpt-6-luna",
          "runner": "benchmark_agent",
          "config": "openai-gpt-6-luna-low",
          "verified_publish": true,
          "levels_completed": 0,
          "actions": 110,
          "resets": 4,
          "state": "NOT_FINISHED",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            110,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            22,
            123,
            73,
            84,
            96,
            192,
            186
          ]
        },
        {
          "session_id": "71179462-28af-45eb-96cd-18d254554fc0",
          "game_id": "ls20",
          "game_version": "9607627b",
          "published_at": "2026-09-16T19:23:44.057789Z",
          "score": 0,
          "model": "gpt-6-luna",
          "runner": "benchmark_agent",
          "config": "openai-gpt-6-luna-none-provider-adapter",
          "verified_publish": true,
          "levels_completed": 0,
          "actions": 110,
          "resets": 1,
          "state": "NOT_FINISHED",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            110,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            22,
            123,
            73,
            84,
            96,
            192,
            186
          ]
        },
        {
          "session_id": "4760dcda-56f4-4420-9d95-9c91098007e6",
          "game_id": "ls20",
          "game_version": "9607627b",
          "published_at": "2026-09-16T19:23:35.772227Z",
          "score": 0,
          "model": "gpt-6-luna",
          "runner": "benchmark_agent",
          "config": "openai-gpt-6-luna-low-provider-adapter",
          "verified_publish": true,
          "levels_completed": 0,
          "actions": 110,
          "resets": 1,
          "state": "NOT_FINISHED",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            110,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            22,
            123,
            73,
            84,
            96,
            192,
            186
          ]
        },
        {
          "session_id": "1f36efa2-3609-4f47-9a6b-ac5522cb91f2",
          "game_id": "ls20",
          "game_version": "9607627b",
          "published_at": "2026-09-16T19:23:29.273689Z",
          "score": 0,
          "model": "gpt-6-luna",
          "runner": "benchmark_agent",
          "config": "openai-gpt-6-luna-none",
          "verified_publish": true,
          "levels_completed": 0,
          "actions": 110,
          "resets": 4,
          "state": "NOT_FINISHED",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            110,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            22,
            123,
            73,
            84,
            96,
            192,
            186
          ]
        },
        {
          "session_id": "0b01848f-1038-4c75-bf63-33ec5918f6a7",
          "game_id": "ls20",
          "game_version": "9607627b",
          "published_at": "2026-09-16T12:15:29.036381Z",
          "score": 35.714285714285715,
          "model": "gpt-6-1-sol",
          "runner": "benchmark_agent",
          "config": "openai-gpt-6-1-sol-max",
          "verified_publish": true,
          "levels_completed": 4,
          "actions": 659,
          "resets": 7,
          "state": "NOT_FINISHED",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            115,
            115,
            115,
            115,
            0,
            0,
            0
          ],
          "level_actions": [
            17,
            59,
            41,
            62,
            480,
            0,
            0
          ],
          "level_baseline_actions": [
            22,
            123,
            73,
            84,
            96,
            192,
            186
          ]
        },
        {
          "session_id": "3692d59b-7f98-4a37-b5f7-04b3476224a2",
          "game_id": "ls20",
          "game_version": "9607627b",
          "published_at": "2026-09-16T09:02:34.418718Z",
          "score": 0,
          "model": "gpt-6-luna",
          "runner": "benchmark_agent",
          "config": "openai-gpt-6-luna-medium-provider-adapter",
          "verified_publish": true,
          "levels_completed": 0,
          "actions": 110,
          "resets": 2,
          "state": "NOT_FINISHED",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            110,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            22,
            123,
            73,
            84,
            96,
            192,
            186
          ]
        },
        {
          "session_id": "137159a3-f27f-4871-b310-11b780af27bf",
          "game_id": "ls20",
          "game_version": "9607627b",
          "published_at": "2026-09-16T08:59:25.45678Z",
          "score": 0,
          "model": "gpt-6-luna",
          "runner": "benchmark_agent",
          "config": "openai-gpt-6-luna-medium",
          "verified_publish": true,
          "levels_completed": 0,
          "actions": 110,
          "resets": 1,
          "state": "NOT_FINISHED",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            110,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            22,
            123,
            73,
            84,
            96,
            192,
            186
          ]
        },
        {
          "session_id": "e8480162-ec66-4227-aa86-dc0338f6b3c4",
          "game_id": "ls20",
          "game_version": "9607627b",
          "published_at": "2026-09-16T08:36:25.290561Z",
          "score": 0,
          "model": "gpt-6-luna",
          "runner": "benchmark_agent",
          "config": "openai-gpt-6-luna-high",
          "verified_publish": true,
          "levels_completed": 0,
          "actions": 110,
          "resets": 4,
          "state": "NOT_FINISHED",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            110,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            22,
            123,
            73,
            84,
            96,
            192,
            186
          ]
        },
        {
          "session_id": "daeb4e38-b5e0-4941-a0b6-08c864cae36a",
          "game_id": "ls20",
          "game_version": "9607627b",
          "published_at": "2026-09-16T08:30:13.153797Z",
          "score": 0,
          "model": "gpt-6-luna",
          "runner": "benchmark_agent",
          "config": "openai-gpt-6-luna-high-provider-adapter",
          "verified_publish": true,
          "levels_completed": 0,
          "actions": 110,
          "resets": 0,
          "state": "NOT_FINISHED",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            110,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            22,
            123,
            73,
            84,
            96,
            192,
            186
          ]
        },
        {
          "session_id": "7fe820e1-6da7-4064-9c96-644f60d44890",
          "game_id": "ls20",
          "game_version": "9607627b",
          "published_at": "2026-09-16T07:52:40.627006Z",
          "score": 0,
          "model": "gpt-6-luna",
          "runner": "benchmark_agent",
          "config": "openai-gpt-6-luna-xhigh-provider-adapter",
          "verified_publish": true,
          "levels_completed": 0,
          "actions": 110,
          "resets": 4,
          "state": "NOT_FINISHED",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            110,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            22,
            123,
            73,
            84,
            96,
            192,
            186
          ]
        },
        {
          "session_id": "417ab6dd-b571-4eb7-98f3-25967abb5814",
          "game_id": "ls20",
          "game_version": "9607627b",
          "published_at": "2026-09-16T07:31:03.32005Z",
          "score": 3.571428571428571,
          "model": "gpt-6-1-sol",
          "runner": "benchmark_agent",
          "config": "openai-gpt-6-1-sol-low",
          "verified_publish": true,
          "levels_completed": 1,
          "actions": 633,
          "resets": 3,
          "state": "GAME_OVER",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            115,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            18,
            615,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            22,
            123,
            73,
            84,
            96,
            192,
            186
          ]
        },
        {
          "session_id": "2ffd8b2e-78e9-462e-bc54-7bb39254ab3e",
          "game_id": "ls20",
          "game_version": "9607627b",
          "published_at": "2026-09-16T07:28:07.794646Z",
          "score": 0,
          "model": "gpt-6-luna",
          "runner": "benchmark_agent",
          "config": "openai-gpt-6-luna-xhigh",
          "verified_publish": true,
          "levels_completed": 0,
          "actions": 110,
          "resets": 6,
          "state": "NOT_FINISHED",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            110,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            22,
            123,
            73,
            84,
            96,
            192,
            186
          ]
        },
        {
          "session_id": "dfdc754c-62da-4f84-8906-7da4b740e960",
          "game_id": "ls20",
          "game_version": "9607627b",
          "published_at": "2026-09-16T07:20:22.197234Z",
          "score": 0,
          "model": "gpt-6-luna",
          "runner": "benchmark_agent",
          "config": "openai-gpt-6-luna-max",
          "verified_publish": true,
          "levels_completed": 0,
          "actions": 110,
          "resets": 7,
          "state": "NOT_FINISHED",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            110,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            22,
            123,
            73,
            84,
            96,
            192,
            186
          ]
        },
        {
          "session_id": "284542fd-6785-4a1c-ae71-1a54eedb11b2",
          "game_id": "ls20",
          "game_version": "9607627b",
          "published_at": "2026-09-16T06:17:58.993436Z",
          "score": 35.714285714285715,
          "model": "gpt-6-1-sol",
          "runner": "benchmark_agent",
          "config": "openai-gpt-6-1-sol-xhigh",
          "verified_publish": true,
          "levels_completed": 4,
          "actions": 676,
          "resets": 8,
          "state": "NOT_FINISHED",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            115,
            115,
            108.75510204081633,
            115,
            0,
            0,
            0
          ],
          "level_actions": [
            15,
            64,
            70,
            47,
            480,
            0,
            0
          ],
          "level_baseline_actions": [
            22,
            123,
            73,
            84,
            96,
            192,
            186
          ]
        },
        {
          "session_id": "ab4a8722-58ac-4f12-abfc-c1fd085b4fe9",
          "game_id": "ls20",
          "game_version": "9607627b",
          "published_at": "2026-09-16T06:08:54.316349Z",
          "score": 3.571428571428571,
          "model": "gpt-6-1-sol",
          "runner": "benchmark_agent",
          "config": "openai-gpt-6-1-sol-medium",
          "verified_publish": true,
          "levels_completed": 1,
          "actions": 632,
          "resets": 4,
          "state": "GAME_OVER",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            115,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            17,
            615,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            22,
            123,
            73,
            84,
            96,
            192,
            186
          ]
        },
        {
          "session_id": "db02dd4d-4c6e-4f81-a858-ecbd605d782c",
          "game_id": "ls20",
          "game_version": "9607627b",
          "published_at": "2026-09-16T05:28:56.715335Z",
          "score": 100,
          "model": "gpt-6-1-sol",
          "runner": "benchmark_agent",
          "config": "openai-gpt-6-1-sol-high-provider-adapter",
          "verified_publish": true,
          "levels_completed": 7,
          "actions": 500,
          "resets": 0,
          "state": "WIN",
          "completed": true,
          "number_of_levels": 0,
          "level_scores": [
            115,
            115,
            115,
            115,
            73.46938775510203,
            115,
            115
          ],
          "level_actions": [
            17,
            99,
            49,
            48,
            112,
            82,
            93
          ],
          "level_baseline_actions": [
            22,
            123,
            73,
            84,
            96,
            192,
            186
          ]
        },
        {
          "session_id": "1e9d9cb9-8eeb-4a79-a671-9d695096c396",
          "game_id": "ls20",
          "game_version": "9607627b",
          "published_at": "2026-09-16T04:55:53.877801Z",
          "score": 19.067225897920604,
          "model": "gpt-6-1-sol",
          "runner": "benchmark_agent",
          "config": "openai-gpt-6-1-sol-high",
          "verified_publish": true,
          "levels_completed": 3,
          "actions": 593,
          "resets": 3,
          "state": "GAME_OVER",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            115,
            115,
            62.96077504725898,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            17,
            64,
            92,
            420,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            22,
            123,
            73,
            84,
            96,
            192,
            186
          ]
        },
        {
          "session_id": "924e2ec3-dd4d-41f9-995e-0317a274cc1c",
          "game_id": "ls20",
          "game_version": "9607627b",
          "published_at": "2026-09-16T04:16:31.687517Z",
          "score": 100,
          "model": "gpt-6-1-sol",
          "runner": "benchmark_agent",
          "config": "openai-gpt-6-1-sol-xhigh-provider-adapter",
          "verified_publish": true,
          "levels_completed": 7,
          "actions": 484,
          "resets": 0,
          "state": "WIN",
          "completed": true,
          "number_of_levels": 0,
          "level_scores": [
            100,
            115,
            115,
            115,
            69.68620037807183,
            115,
            115
          ],
          "level_actions": [
            22,
            99,
            41,
            52,
            115,
            86,
            69
          ],
          "level_baseline_actions": [
            22,
            123,
            73,
            84,
            96,
            192,
            186
          ]
        },
        {
          "session_id": "043d449f-74b4-4326-a84b-21f73bc85294",
          "game_id": "ls20",
          "game_version": "9607627b",
          "published_at": "2026-09-16T04:13:46.236238Z",
          "score": 100,
          "model": "gpt-6-1-sol",
          "runner": "benchmark_agent",
          "config": "openai-gpt-6-1-sol-max-provider-adapter",
          "verified_publish": true,
          "levels_completed": 7,
          "actions": 397,
          "resets": 0,
          "state": "WIN",
          "completed": true,
          "number_of_levels": 0,
          "level_scores": [
            115,
            115,
            115,
            115,
            115,
            115,
            115
          ],
          "level_actions": [
            15,
            45,
            49,
            48,
            77,
            82,
            81
          ],
          "level_baseline_actions": [
            22,
            123,
            73,
            84,
            96,
            192,
            186
          ]
        },
        {
          "session_id": "17c561f7-852b-4c61-8e63-5be985dc0273",
          "game_id": "ls20",
          "game_version": "9607627b",
          "published_at": "2026-09-16T02:22:45.407332Z",
          "score": 0,
          "model": "gpt-6-luna",
          "runner": "benchmark_agent",
          "config": "openai-gpt-6-luna-max-provider-adapter",
          "verified_publish": true,
          "levels_completed": 0,
          "actions": 110,
          "resets": 1,
          "state": "NOT_FINISHED",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            110,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            22,
            123,
            73,
            84,
            96,
            192,
            186
          ]
        },
        {
          "session_id": "4a8cca1e-ce2f-46a2-be27-f4f449323491",
          "game_id": "ls20",
          "game_version": "9607627b",
          "published_at": "2026-09-12T07:44:15.857022Z",
          "score": 1.411078717201166,
          "model": "gemini-3.8-flash",
          "runner": "benchmark_agent",
          "config": "google-gemini-3-8-flash-low",
          "verified_publish": true,
          "levels_completed": 1,
          "actions": 650,
          "resets": 4,
          "state": "GAME_OVER",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            39.51020408163265,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            35,
            615,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            22,
            123,
            73,
            84,
            96,
            192,
            186
          ]
        },
        {
          "session_id": "093b8d55-09f1-40b8-a980-47a15d5af3e2",
          "game_id": "ls20",
          "game_version": "9607627b",
          "published_at": "2026-09-09T19:00:17.971049Z",
          "score": 10.714285714285714,
          "model": "gemini-3.8-flash",
          "runner": "benchmark_agent",
          "config": "google-gemini-3-8-flash-high",
          "verified_publish": true,
          "levels_completed": 2,
          "actions": 271,
          "resets": 4,
          "state": "NOT_FINISHED",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            109.75056689342404,
            115,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            21,
            104,
            146,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            22,
            123,
            73,
            84,
            96,
            192,
            186
          ]
        },
        {
          "session_id": "d6b29990-38a7-4c5b-b13c-c559121fb55b",
          "game_id": "ls20",
          "game_version": "9607627b",
          "published_at": "2026-09-09T17:38:38.21437Z",
          "score": 0.4355181225929525,
          "model": "gemini-3.8-flash",
          "runner": "benchmark_agent",
          "config": "google-gemini-3-8-flash-medium",
          "verified_publish": true,
          "levels_completed": 1,
          "actions": 207,
          "resets": 0,
          "state": "NOT_FINISHED",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            12.19450743260267,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            63,
            144,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            22,
            123,
            73,
            84,
            96,
            192,
            186
          ]
        },
        {
          "session_id": "e1ec9bbf-6711-412b-86da-1077467415fb",
          "game_id": "ls20",
          "game_version": "9607627b",
          "published_at": "2026-08-11T07:37:39.332906Z",
          "score": 0,
          "model": "grok-4.6",
          "runner": "benchmark_agent",
          "config": "xai-grok-4-6-xhigh",
          "verified_publish": true,
          "levels_completed": 0,
          "actions": 110,
          "resets": 1,
          "state": "NOT_FINISHED",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            110,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            22,
            123,
            73,
            84,
            96,
            192,
            186
          ]
        },
        {
          "session_id": "678595a5-808c-4e3e-9074-1c5d2cbf1b23",
          "game_id": "ls20",
          "game_version": "9607627b",
          "published_at": "2026-07-21T09:12:52.169475Z",
          "score": 5.362414730598121,
          "model": "claude-opus-5",
          "runner": "benchmark_agent",
          "config": "anthropic-claude-opus-5-high",
          "verified_publish": true,
          "levels_completed": 2,
          "actions": 594,
          "resets": 7,
          "state": "NOT_FINISHED",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            77.44,
            36.3538062283737,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            25,
            204,
            365,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            22,
            123,
            73,
            84,
            96,
            192,
            186
          ]
        },
        {
          "session_id": "58c9ab03-43de-479c-9946-a77d9755a2b0",
          "game_id": "ls20",
          "game_version": "9607627b",
          "published_at": "2026-06-24T10:55:20.329443Z",
          "score": 0,
          "model": "gpt-5-6-luna",
          "runner": "benchmark_agent",
          "config": "openai-gpt-5-6-luna-low",
          "verified_publish": true,
          "levels_completed": 0,
          "actions": 110,
          "resets": 3,
          "state": "NOT_FINISHED",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            110,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            22,
            123,
            73,
            84,
            96,
            192,
            186
          ]
        },
        {
          "session_id": "067b2629-d848-435a-9324-d441b2acc9d5",
          "game_id": "ls20",
          "game_version": "9607627b",
          "published_at": "2026-06-24T10:33:02.197349Z",
          "score": 0,
          "model": "gpt-5-6-luna",
          "runner": "benchmark_agent",
          "config": "openai-gpt-5-6-luna-medium",
          "verified_publish": true,
          "levels_completed": 0,
          "actions": 110,
          "resets": 2,
          "state": "NOT_FINISHED",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            110,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            22,
            123,
            73,
            84,
            96,
            192,
            186
          ]
        },
        {
          "session_id": "56e62036-590f-4589-a2b7-03a3e89e6ff7",
          "game_id": "ls20",
          "game_version": "9607627b",
          "published_at": "2026-06-24T03:54:36.36999Z",
          "score": 0,
          "model": "gpt-5-6-luna",
          "runner": "benchmark_agent",
          "config": "openai-gpt-5-6-luna-high",
          "verified_publish": true,
          "levels_completed": 0,
          "actions": 110,
          "resets": 17,
          "state": "NOT_FINISHED",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            110,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            22,
            123,
            73,
            84,
            96,
            192,
            186
          ]
        },
        {
          "session_id": "c7198793-5c97-4099-acc8-bfc2488b2c89",
          "game_id": "ls20",
          "game_version": "9607627b",
          "published_at": "2026-06-24T01:49:31.518506Z",
          "score": 0,
          "model": "gpt-5-6-luna",
          "runner": "benchmark_agent",
          "config": "openai-gpt-5-6-luna-xhigh",
          "verified_publish": true,
          "levels_completed": 0,
          "actions": 110,
          "resets": 21,
          "state": "NOT_FINISHED",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            110,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            22,
            123,
            73,
            84,
            96,
            192,
            186
          ]
        },
        {
          "session_id": "5cbced8c-2974-4282-90ff-da348ee61d9c",
          "game_id": "ls20",
          "game_version": "9607627b",
          "published_at": "2026-06-23T01:42:34.017996Z",
          "score": 0,
          "model": "gpt-5-6-luna",
          "runner": "benchmark_agent",
          "config": "openai-gpt-5-6-luna-max",
          "verified_publish": true,
          "levels_completed": 0,
          "actions": 110,
          "resets": 14,
          "state": "NOT_FINISHED",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            110,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            22,
            123,
            73,
            84,
            96,
            192,
            186
          ]
        },
        {
          "session_id": "9d2a5f8e-8a56-4a1f-b205-66017ca68079",
          "game_id": "ls20",
          "game_version": "9607627b",
          "published_at": "2026-06-20T12:00:48.094354Z",
          "score": 0,
          "model": "gpt-5-6-terra",
          "runner": "benchmark_agent",
          "config": "openai-gpt-5-6-terra-low",
          "verified_publish": true,
          "levels_completed": 0,
          "actions": 110,
          "resets": 0,
          "state": "NOT_FINISHED",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            110,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            22,
            123,
            73,
            84,
            96,
            192,
            186
          ]
        },
        {
          "session_id": "6e425b4d-a708-4110-89a5-b74f31575751",
          "game_id": "ls20",
          "game_version": "9607627b",
          "published_at": "2026-06-20T09:26:08.822191Z",
          "score": 0,
          "model": "gpt-5-6-terra",
          "runner": "benchmark_agent",
          "config": "openai-gpt-5-6-terra-medium",
          "verified_publish": true,
          "levels_completed": 0,
          "actions": 110,
          "resets": 0,
          "state": "NOT_FINISHED",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            110,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            22,
            123,
            73,
            84,
            96,
            192,
            186
          ]
        },
        {
          "session_id": "d0915150-3a23-4bc2-bda0-d2b37d8cc461",
          "game_id": "ls20",
          "game_version": "9607627b",
          "published_at": "2026-06-20T08:59:00.373078Z",
          "score": 0,
          "model": "gpt-5-6-terra",
          "runner": "benchmark_agent",
          "config": "openai-gpt-5-6-terra-high",
          "verified_publish": true,
          "levels_completed": 0,
          "actions": 110,
          "resets": 0,
          "state": "NOT_FINISHED",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            110,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            22,
            123,
            73,
            84,
            96,
            192,
            186
          ]
        },
        {
          "session_id": "10b7991b-2ef9-4c3e-b158-902cf8e95fd5",
          "game_id": "ls20",
          "game_version": "9607627b",
          "published_at": "2026-06-20T08:39:31.352924Z",
          "score": 1.6880580357142858,
          "model": "gpt-5-6-terra",
          "runner": "benchmark_agent",
          "config": "openai-gpt-5-6-terra-xhigh",
          "verified_publish": true,
          "levels_completed": 1,
          "actions": 647,
          "resets": 14,
          "state": "GAME_OVER",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            47.265625,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            32,
            615,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            22,
            123,
            73,
            84,
            96,
            192,
            186
          ]
        },
        {
          "session_id": "1e097441-89bf-44f2-adb9-cb3d4e3a4ad3",
          "game_id": "ls20",
          "game_version": "9607627b",
          "published_at": "2026-06-18T21:29:46.578152Z",
          "score": 0,
          "model": "gpt-5-6-terra",
          "runner": "benchmark_agent",
          "config": "openai-gpt-5-6-terra-max",
          "verified_publish": true,
          "levels_completed": 0,
          "actions": 110,
          "resets": 6,
          "state": "NOT_FINISHED",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            110,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            22,
            123,
            73,
            84,
            96,
            192,
            186
          ]
        },
        {
          "session_id": "989a45d3-5f0f-4280-9073-b94432dbc09d",
          "game_id": "ls20",
          "game_version": "9607627b",
          "published_at": "2026-06-13T13:17:51.135874Z",
          "score": 3.571428571428571,
          "model": "gpt-5-6-sol",
          "runner": "benchmark_agent",
          "config": "openai-gpt-5-6-sol-max",
          "verified_publish": true,
          "levels_completed": 1,
          "actions": 632,
          "resets": 16,
          "state": "GAME_OVER",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            115,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            17,
            615,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            22,
            123,
            73,
            84,
            96,
            192,
            186
          ]
        },
        {
          "session_id": "81bc89ff-fe24-4aab-adce-6394b27ebfee",
          "game_id": "ls20",
          "game_version": "9607627b",
          "published_at": "2026-06-12T09:03:49.308897Z",
          "score": 1.411078717201166,
          "model": "gpt-5-6-sol",
          "runner": "benchmark_agent",
          "config": "openai-gpt-5-6-sol-high",
          "verified_publish": true,
          "levels_completed": 1,
          "actions": 650,
          "resets": 15,
          "state": "GAME_OVER",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            39.51020408163265,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            35,
            615,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            22,
            123,
            73,
            84,
            96,
            192,
            186
          ]
        },
        {
          "session_id": "83543d22-8e1e-439a-8809-129ff1d9ebcd",
          "game_id": "ls20",
          "game_version": "9607627b",
          "published_at": "2026-06-12T00:31:26.650959Z",
          "score": 3.652934306192654,
          "model": "gpt-5-6-sol",
          "runner": "benchmark_agent",
          "config": "openai-gpt-5-6-sol-xhigh",
          "verified_publish": true,
          "levels_completed": 2,
          "actions": 567,
          "resets": 27,
          "state": "NOT_FINISHED",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            77.44,
            12.421080286697155,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            25,
            349,
            193,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            22,
            123,
            73,
            84,
            96,
            192,
            186
          ]
        },
        {
          "session_id": "4219a5ea-b1f3-44ca-af98-ee75f111313c",
          "game_id": "ls20",
          "game_version": "9607627b",
          "published_at": "2026-06-10T08:48:36.244303Z",
          "score": 0,
          "model": "gpt-5-6-sol",
          "runner": "benchmark_agent",
          "config": "openai-gpt-5-6-sol-medium",
          "verified_publish": true,
          "levels_completed": 0,
          "actions": 110,
          "resets": 0,
          "state": "NOT_FINISHED",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            110,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            22,
            123,
            73,
            84,
            96,
            192,
            186
          ]
        },
        {
          "session_id": "b16d4919-c689-44bf-84f9-fcfd48fb4c74",
          "game_id": "ls20",
          "game_version": "9607627b",
          "published_at": "2026-06-09T22:13:03.636822Z",
          "score": 0,
          "model": "gpt-5-6-sol",
          "runner": "benchmark_agent",
          "config": "openai-gpt-5-6-sol-low",
          "verified_publish": true,
          "levels_completed": 0,
          "actions": 110,
          "resets": 0,
          "state": "NOT_FINISHED",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            110,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            22,
            123,
            73,
            84,
            96,
            192,
            186
          ]
        },
        {
          "session_id": "8314341d-c2e5-4b75-af8a-a085eddd8165",
          "game_id": "ls20",
          "game_version": "9607627b",
          "published_at": "2026-05-31T20:49:31.346372Z",
          "score": 2.055376252760319,
          "model": "claude-opus-4-8",
          "runner": "benchmark_agent",
          "config": "anthropic-opus-4-8-high",
          "verified_publish": true,
          "levels_completed": 1,
          "actions": 644,
          "resets": 14,
          "state": "GAME_OVER",
          "completed": false,
          "number_of_levels": 7,
          "level_scores": [
            57.550535077288934,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            29,
            615,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            22,
            123,
            73,
            84,
            96,
            192,
            186
          ]
        },
        {
          "session_id": "21578c6c-b05b-4fb2-ac87-d60e679b2698",
          "game_id": "ls20",
          "game_version": "9607627b",
          "published_at": "2026-05-01T05:07:39.705991Z",
          "score": 0,
          "model": "claude-opus-4-7",
          "runner": "benchmark_agent",
          "config": "anthropic-opus-4-7-high",
          "verified_publish": true,
          "levels_completed": 0,
          "actions": 110,
          "resets": 3,
          "state": "NOT_FINISHED",
          "completed": false,
          "number_of_levels": 7,
          "level_scores": [
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            110,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            22,
            123,
            73,
            84,
            96,
            192,
            186
          ]
        },
        {
          "session_id": "d4472541-4670-4c34-92f4-354af3d82a87",
          "game_id": "ls20",
          "game_version": "9607627b",
          "published_at": "2026-04-30T23:25:08.552342Z",
          "score": 0,
          "model": "gpt-5.5-2026-04-23",
          "runner": "benchmark_agent",
          "config": "openai-gpt-5-5-2026-04-23-high",
          "verified_publish": true,
          "levels_completed": 0,
          "actions": 110,
          "resets": 4,
          "state": "NOT_FINISHED",
          "completed": false,
          "number_of_levels": 7,
          "level_scores": [
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            110,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            22,
            123,
            73,
            84,
            96,
            192,
            186
          ]
        },
        {
          "session_id": "7b705f94-c41d-476f-8d39-12ca71941cf7",
          "game_id": "ls20",
          "game_version": "9607627b",
          "published_at": "2026-04-14T17:43:39.757365Z",
          "score": 0,
          "model": "grok-4.20-beta-0309-reasoning",
          "runner": "benchmark_agent",
          "config": "xai-grok-4-20-beta-0309-reasoning",
          "verified_publish": true,
          "levels_completed": 0,
          "actions": 110,
          "resets": 6,
          "state": "NOT_FINISHED",
          "completed": false,
          "number_of_levels": 7,
          "level_scores": [
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            110,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            22,
            123,
            73,
            84,
            96,
            192,
            186
          ]
        },
        {
          "session_id": "1f8ca30f-d063-460f-9edd-d02f34eecb31",
          "game_id": "ls20",
          "game_version": "9607627b",
          "published_at": "2026-04-14T17:43:39.48218Z",
          "score": 0,
          "model": "gpt-5.4-2026-03-05",
          "runner": "benchmark_agent",
          "config": "openai-gpt-5-4-2026-03-05-high",
          "verified_publish": true,
          "levels_completed": 0,
          "actions": 110,
          "resets": 2,
          "state": "NOT_FINISHED",
          "completed": false,
          "number_of_levels": 7,
          "level_scores": [
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            110,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            22,
            123,
            73,
            84,
            96,
            192,
            186
          ]
        },
        {
          "session_id": "2b4425c2-a622-461e-9013-18a20022200a",
          "game_id": "ls20",
          "game_version": "9607627b",
          "published_at": "2026-04-14T17:43:38.042593Z",
          "score": 0,
          "model": "gemini-3.1-pro-preview",
          "runner": "benchmark_agent",
          "config": "google-gemini-3-1-pro-preview",
          "verified_publish": true,
          "levels_completed": 0,
          "actions": 110,
          "resets": 0,
          "state": "NOT_FINISHED",
          "completed": false,
          "number_of_levels": 7,
          "level_scores": [
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            110,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            22,
            123,
            73,
            84,
            96,
            192,
            186
          ]
        },
        {
          "session_id": "a8e7de3c-ca5c-4822-8dba-2b3d68b34cbf",
          "game_id": "ls20",
          "game_version": "9607627b",
          "published_at": "2026-04-14T17:43:37.762615Z",
          "score": 0,
          "model": "claude-opus-4-6",
          "runner": "benchmark_agent",
          "config": "anthropic-opus-4-6-max-effort",
          "verified_publish": true,
          "levels_completed": 0,
          "actions": 110,
          "resets": 5,
          "state": "NOT_FINISHED",
          "completed": false,
          "number_of_levels": 7,
          "level_scores": [
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            110,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            22,
            123,
            73,
            84,
            96,
            192,
            186
          ]
        }
      ]
    },
    {
      "game": "bp35",
      "source": "https://arcprize.org/tasks/bp35",
      "runs": [
        {
          "session_id": "9ab25d1b-4ad8-4469-84b7-af51bf36b59e",
          "game_id": "bp35",
          "game_version": "0a0ad940",
          "published_at": "2026-09-25T06:38:22.707768Z",
          "score": 0,
          "model": "gpt-6-sol",
          "runner": "benchmark_agent",
          "config": "openai-gpt-6-sol-none-provider-adapter",
          "verified_publish": true,
          "levels_completed": 0,
          "actions": 105,
          "resets": 5,
          "state": "GAME_OVER",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            0,
            0,
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            105,
            0,
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            21,
            48,
            44,
            38,
            33,
            87,
            86,
            131,
            163
          ]
        },
        {
          "session_id": "abb4ebaf-af23-47bd-9dd5-2055a00c4eca",
          "game_id": "bp35",
          "game_version": "0a0ad940",
          "published_at": "2026-09-25T06:22:40.445251Z",
          "score": 0,
          "model": "gpt-6-sol",
          "runner": "benchmark_agent",
          "config": "openai-gpt-6-sol-none",
          "verified_publish": true,
          "levels_completed": 0,
          "actions": 105,
          "resets": 2,
          "state": "GAME_OVER",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            0,
            0,
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            105,
            0,
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            21,
            48,
            44,
            38,
            33,
            87,
            86,
            131,
            163
          ]
        },
        {
          "session_id": "c5116198-7baf-49c4-93cd-f1805ece11ae",
          "game_id": "bp35",
          "game_version": "0a0ad940",
          "published_at": "2026-09-25T06:12:45.031452Z",
          "score": 0,
          "model": "gpt-6-sol",
          "runner": "benchmark_agent",
          "config": "openai-gpt-6-sol-low-provider-adapter",
          "verified_publish": true,
          "levels_completed": 0,
          "actions": 105,
          "resets": 2,
          "state": "GAME_OVER",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            0,
            0,
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            105,
            0,
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            21,
            48,
            44,
            38,
            33,
            87,
            86,
            131,
            163
          ]
        },
        {
          "session_id": "a30e6ed2-58cc-4d3f-89e7-db4de753b8a7",
          "game_id": "bp35",
          "game_version": "0a0ad940",
          "published_at": "2026-09-25T05:54:41.772438Z",
          "score": 0,
          "model": "gpt-6-sol",
          "runner": "benchmark_agent",
          "config": "openai-gpt-6-sol-low",
          "verified_publish": true,
          "levels_completed": 0,
          "actions": 105,
          "resets": 1,
          "state": "GAME_OVER",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            0,
            0,
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            105,
            0,
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            21,
            48,
            44,
            38,
            33,
            87,
            86,
            131,
            163
          ]
        },
        {
          "session_id": "7abf5685-ff10-41d4-89ba-24f230c353b0",
          "game_id": "bp35",
          "game_version": "0a0ad940",
          "published_at": "2026-09-25T05:44:40.043074Z",
          "score": 0,
          "model": "gpt-6-sol",
          "runner": "benchmark_agent",
          "config": "openai-gpt-6-sol-medium-provider-adapter",
          "verified_publish": true,
          "levels_completed": 0,
          "actions": 105,
          "resets": 1,
          "state": "GAME_OVER",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            0,
            0,
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            105,
            0,
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            21,
            48,
            44,
            38,
            33,
            87,
            86,
            131,
            163
          ]
        },
        {
          "session_id": "a0874076-5e6b-440f-a585-d25fef904c27",
          "game_id": "bp35",
          "game_version": "0a0ad940",
          "published_at": "2026-09-25T05:21:46.741246Z",
          "score": 0,
          "model": "gpt-6-sol",
          "runner": "benchmark_agent",
          "config": "openai-gpt-6-sol-medium",
          "verified_publish": true,
          "levels_completed": 0,
          "actions": 105,
          "resets": 2,
          "state": "GAME_OVER",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            0,
            0,
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            105,
            0,
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            21,
            48,
            44,
            38,
            33,
            87,
            86,
            131,
            163
          ]
        },
        {
          "session_id": "be053640-29dc-4edd-9c83-a322cff34c3c",
          "game_id": "bp35",
          "game_version": "0a0ad940",
          "published_at": "2026-09-25T04:58:31.096362Z",
          "score": 0.4436396559529199,
          "model": "gpt-6-sol",
          "runner": "benchmark_agent",
          "config": "openai-gpt-6-sol-xhigh-provider-adapter",
          "verified_publish": true,
          "levels_completed": 1,
          "actions": 287,
          "resets": 19,
          "state": "GAME_OVER",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            19.963784517881393,
            0,
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            47,
            240,
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            21,
            48,
            44,
            38,
            33,
            87,
            86,
            131,
            163
          ]
        },
        {
          "session_id": "e3418cbc-80e3-40dd-8e30-31e814677d94",
          "game_id": "bp35",
          "game_version": "0a0ad940",
          "published_at": "2026-09-25T04:55:34.762833Z",
          "score": 0,
          "model": "gpt-6-sol",
          "runner": "benchmark_agent",
          "config": "openai-gpt-6-sol-high-provider-adapter",
          "verified_publish": true,
          "levels_completed": 0,
          "actions": 105,
          "resets": 2,
          "state": "GAME_OVER",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            0,
            0,
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            105,
            0,
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            21,
            48,
            44,
            38,
            33,
            87,
            86,
            131,
            163
          ]
        },
        {
          "session_id": "9eead73b-2ca6-4c4f-b854-1cf5dbd89349",
          "game_id": "bp35",
          "game_version": "0a0ad940",
          "published_at": "2026-09-25T04:43:36.349164Z",
          "score": 0,
          "model": "gpt-6-sol",
          "runner": "benchmark_agent",
          "config": "openai-gpt-6-sol-max",
          "verified_publish": true,
          "levels_completed": 0,
          "actions": 105,
          "resets": 1,
          "state": "GAME_OVER",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            0,
            0,
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            105,
            0,
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            21,
            48,
            44,
            38,
            33,
            87,
            86,
            131,
            163
          ]
        },
        {
          "session_id": "57b97296-480a-4b3e-a14f-9b4c6dfb0c56",
          "game_id": "bp35",
          "game_version": "0a0ad940",
          "published_at": "2026-09-25T04:31:32.880523Z",
          "score": 0,
          "model": "gpt-6-sol",
          "runner": "benchmark_agent",
          "config": "openai-gpt-6-sol-high",
          "verified_publish": true,
          "levels_completed": 0,
          "actions": 105,
          "resets": 1,
          "state": "GAME_OVER",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            0,
            0,
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            105,
            0,
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            21,
            48,
            44,
            38,
            33,
            87,
            86,
            131,
            163
          ]
        },
        {
          "session_id": "4bf62fb3-e176-423b-8399-d768a66558be",
          "game_id": "bp35",
          "game_version": "0a0ad940",
          "published_at": "2026-09-25T04:12:20.172301Z",
          "score": 0,
          "model": "gpt-6-sol",
          "runner": "benchmark_agent",
          "config": "openai-gpt-6-sol-xhigh",
          "verified_publish": true,
          "levels_completed": 0,
          "actions": 105,
          "resets": 1,
          "state": "GAME_OVER",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            0,
            0,
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            105,
            0,
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            21,
            48,
            44,
            38,
            33,
            87,
            86,
            131,
            163
          ]
        },
        {
          "session_id": "9bdd25c1-ca1e-4fd0-aaae-6805f02398f4",
          "game_id": "bp35",
          "game_version": "0a0ad940",
          "published_at": "2026-09-21T08:23:15.495566Z",
          "score": 0.7158509861212563,
          "model": "gemini-3.8-flash",
          "runner": "benchmark_agent",
          "config": "google-gemini-3-8-flash-low-provider-adapter",
          "verified_publish": true,
          "levels_completed": 1,
          "actions": 277,
          "resets": 15,
          "state": "GAME_OVER",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            32.213294375456535,
            0,
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            37,
            240,
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            21,
            48,
            44,
            38,
            33,
            87,
            86,
            131,
            163
          ]
        },
        {
          "session_id": "79420967-5ab4-4591-93a1-f401bc3e49fb",
          "game_id": "bp35",
          "game_version": "0a0ad940",
          "published_at": "2026-09-20T03:52:00.553686Z",
          "score": 2.2222222222222223,
          "model": "gemini-3.8-flash",
          "runner": "benchmark_agent",
          "config": "google-gemini-3-8-flash-medium-provider-adapter",
          "verified_publish": true,
          "levels_completed": 1,
          "actions": 257,
          "resets": 30,
          "state": "GAME_OVER",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            115,
            0,
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            17,
            240,
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            21,
            48,
            44,
            38,
            33,
            87,
            86,
            131,
            163
          ]
        },
        {
          "session_id": "2bad2c09-ef50-420e-818c-8de40aa133d0",
          "game_id": "bp35",
          "game_version": "0a0ad940",
          "published_at": "2026-09-17T19:08:54.150809Z",
          "score": 0.11090991398822997,
          "model": "gemini-3.8-flash",
          "runner": "benchmark_agent",
          "config": "google-gemini-3-8-flash-high-provider-adapter",
          "verified_publish": true,
          "levels_completed": 1,
          "actions": 334,
          "resets": 14,
          "state": "GAME_OVER",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            4.990946129470348,
            0,
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            94,
            240,
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            21,
            48,
            44,
            38,
            33,
            87,
            86,
            131,
            163
          ]
        },
        {
          "session_id": "fbf6fbdc-c957-44f0-9383-d85155e504b7",
          "game_id": "bp35",
          "game_version": "0a0ad940",
          "published_at": "2026-09-17T01:11:36.707047Z",
          "score": 99.05930069091875,
          "model": "gpt-6-1-sol",
          "runner": "benchmark_agent",
          "config": "openai-gpt-6-1-sol-high-provider-adapter",
          "verified_publish": true,
          "levels_completed": 9,
          "actions": 584,
          "resets": 12,
          "state": "WIN",
          "completed": true,
          "number_of_levels": 0,
          "level_scores": [
            115,
            104.30058850158441,
            115,
            115,
            53.77777777777777,
            50.02974419988102,
            115,
            115,
            115
          ],
          "level_actions": [
            18,
            47,
            36,
            27,
            45,
            123,
            67,
            83,
            138
          ],
          "level_baseline_actions": [
            21,
            48,
            44,
            38,
            33,
            87,
            86,
            131,
            163
          ]
        },
        {
          "session_id": "1b3d1f8c-61a0-41e8-a14a-561fafffc1f9",
          "game_id": "bp35",
          "game_version": "0a0ad940",
          "published_at": "2026-09-16T19:26:26.8691Z",
          "score": 0,
          "model": "gpt-6-luna",
          "runner": "benchmark_agent",
          "config": "openai-gpt-6-luna-low-provider-adapter",
          "verified_publish": true,
          "levels_completed": 0,
          "actions": 105,
          "resets": 3,
          "state": "GAME_OVER",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            0,
            0,
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            105,
            0,
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            21,
            48,
            44,
            38,
            33,
            87,
            86,
            131,
            163
          ]
        },
        {
          "session_id": "b425b34b-115f-4d8e-b250-27b2006f4a58",
          "game_id": "bp35",
          "game_version": "0a0ad940",
          "published_at": "2026-09-16T19:25:34.772995Z",
          "score": 0,
          "model": "gpt-6-luna",
          "runner": "benchmark_agent",
          "config": "openai-gpt-6-luna-none-provider-adapter",
          "verified_publish": true,
          "levels_completed": 0,
          "actions": 105,
          "resets": 1,
          "state": "GAME_OVER",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            0,
            0,
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            105,
            0,
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            21,
            48,
            44,
            38,
            33,
            87,
            86,
            131,
            163
          ]
        },
        {
          "session_id": "b87aaa8b-b26b-4d50-810f-21efe1314724",
          "game_id": "bp35",
          "game_version": "0a0ad940",
          "published_at": "2026-09-16T19:25:14.83579Z",
          "score": 0,
          "model": "gpt-6-luna",
          "runner": "benchmark_agent",
          "config": "openai-gpt-6-luna-low",
          "verified_publish": true,
          "levels_completed": 0,
          "actions": 105,
          "resets": 2,
          "state": "GAME_OVER",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            0,
            0,
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            105,
            0,
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            21,
            48,
            44,
            38,
            33,
            87,
            86,
            131,
            163
          ]
        },
        {
          "session_id": "a20c7a16-7294-467e-9e17-7860582ca379",
          "game_id": "bp35",
          "game_version": "0a0ad940",
          "published_at": "2026-09-16T19:22:40.222342Z",
          "score": 0,
          "model": "gpt-6-luna",
          "runner": "benchmark_agent",
          "config": "openai-gpt-6-luna-none",
          "verified_publish": true,
          "levels_completed": 0,
          "actions": 105,
          "resets": 3,
          "state": "GAME_OVER",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            0,
            0,
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            105,
            0,
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            21,
            48,
            44,
            38,
            33,
            87,
            86,
            131,
            163
          ]
        },
        {
          "session_id": "4cf94df3-88ed-45cd-b6d8-a1d3c8f6ec5b",
          "game_id": "bp35",
          "game_version": "0a0ad940",
          "published_at": "2026-09-16T09:08:38.027414Z",
          "score": 0,
          "model": "gpt-6-luna",
          "runner": "benchmark_agent",
          "config": "openai-gpt-6-luna-medium-provider-adapter",
          "verified_publish": true,
          "levels_completed": 0,
          "actions": 105,
          "resets": 3,
          "state": "GAME_OVER",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            0,
            0,
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            105,
            0,
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            21,
            48,
            44,
            38,
            33,
            87,
            86,
            131,
            163
          ]
        },
        {
          "session_id": "80335752-540d-42b0-93ab-b4b179591182",
          "game_id": "bp35",
          "game_version": "0a0ad940",
          "published_at": "2026-09-16T08:58:12.256465Z",
          "score": 0,
          "model": "gpt-6-luna",
          "runner": "benchmark_agent",
          "config": "openai-gpt-6-luna-medium",
          "verified_publish": true,
          "levels_completed": 0,
          "actions": 105,
          "resets": 3,
          "state": "GAME_OVER",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            0,
            0,
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            105,
            0,
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            21,
            48,
            44,
            38,
            33,
            87,
            86,
            131,
            163
          ]
        },
        {
          "session_id": "7afa49d6-eb6a-45d7-b562-43bf8e3270a3",
          "game_id": "bp35",
          "game_version": "0a0ad940",
          "published_at": "2026-09-16T08:44:49.898095Z",
          "score": 0,
          "model": "gpt-6-luna",
          "runner": "benchmark_agent",
          "config": "openai-gpt-6-luna-high-provider-adapter",
          "verified_publish": true,
          "levels_completed": 0,
          "actions": 105,
          "resets": 3,
          "state": "GAME_OVER",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            0,
            0,
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            105,
            0,
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            21,
            48,
            44,
            38,
            33,
            87,
            86,
            131,
            163
          ]
        },
        {
          "session_id": "334605df-934c-4c7c-8c39-fb12959b8383",
          "game_id": "bp35",
          "game_version": "0a0ad940",
          "published_at": "2026-09-16T08:40:09.122746Z",
          "score": 0,
          "model": "gpt-6-luna",
          "runner": "benchmark_agent",
          "config": "openai-gpt-6-luna-high",
          "verified_publish": true,
          "levels_completed": 0,
          "actions": 105,
          "resets": 2,
          "state": "GAME_OVER",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            0,
            0,
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            105,
            0,
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            21,
            48,
            44,
            38,
            33,
            87,
            86,
            131,
            163
          ]
        },
        {
          "session_id": "00e5e616-f93b-464c-9bd8-0ff866e21c01",
          "game_id": "bp35",
          "game_version": "0a0ad940",
          "published_at": "2026-09-16T07:54:14.733702Z",
          "score": 0,
          "model": "gpt-6-luna",
          "runner": "benchmark_agent",
          "config": "openai-gpt-6-luna-xhigh-provider-adapter",
          "verified_publish": true,
          "levels_completed": 0,
          "actions": 105,
          "resets": 1,
          "state": "GAME_OVER",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            0,
            0,
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            105,
            0,
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            21,
            48,
            44,
            38,
            33,
            87,
            86,
            131,
            163
          ]
        },
        {
          "session_id": "165446d1-6c02-4158-a1f1-3185425c3dab",
          "game_id": "bp35",
          "game_version": "0a0ad940",
          "published_at": "2026-09-16T07:32:53.341372Z",
          "score": 0,
          "model": "gpt-6-luna",
          "runner": "benchmark_agent",
          "config": "openai-gpt-6-luna-xhigh",
          "verified_publish": true,
          "levels_completed": 0,
          "actions": 105,
          "resets": 2,
          "state": "GAME_OVER",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            0,
            0,
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            105,
            0,
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            21,
            48,
            44,
            38,
            33,
            87,
            86,
            131,
            163
          ]
        },
        {
          "session_id": "e305a326-64fc-430a-8e82-279c2100f283",
          "game_id": "bp35",
          "game_version": "0a0ad940",
          "published_at": "2026-09-16T07:25:19.514618Z",
          "score": 0,
          "model": "gpt-6-luna",
          "runner": "benchmark_agent",
          "config": "openai-gpt-6-luna-max",
          "verified_publish": true,
          "levels_completed": 0,
          "actions": 105,
          "resets": 2,
          "state": "GAME_OVER",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            0,
            0,
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            105,
            0,
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            21,
            48,
            44,
            38,
            33,
            87,
            86,
            131,
            163
          ]
        },
        {
          "session_id": "0ceba43d-8f4b-4793-af92-878620c46ff0",
          "game_id": "bp35",
          "game_version": "0a0ad940",
          "published_at": "2026-09-16T07:23:02.884576Z",
          "score": 0,
          "model": "gpt-6-luna",
          "runner": "benchmark_agent",
          "config": "openai-gpt-6-luna-max-provider-adapter",
          "verified_publish": true,
          "levels_completed": 0,
          "actions": 105,
          "resets": 4,
          "state": "GAME_OVER",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            0,
            0,
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            105,
            0,
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            21,
            48,
            44,
            38,
            33,
            87,
            86,
            131,
            163
          ]
        },
        {
          "session_id": "b496a0c5-98a0-4a6a-9e63-faf04e1da581",
          "game_id": "bp35",
          "game_version": "0a0ad940",
          "published_at": "2026-09-16T06:58:46.557509Z",
          "score": 1.019771071800208,
          "model": "gpt-6-1-sol",
          "runner": "benchmark_agent",
          "config": "openai-gpt-6-1-sol-low",
          "verified_publish": true,
          "levels_completed": 1,
          "actions": 271,
          "resets": 66,
          "state": "GAME_OVER",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            45.889698231009355,
            0,
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            31,
            240,
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            21,
            48,
            44,
            38,
            33,
            87,
            86,
            131,
            163
          ]
        },
        {
          "session_id": "b14d28ab-b183-4f9f-98d5-8a52a0a915b2",
          "game_id": "bp35",
          "game_version": "0a0ad940",
          "published_at": "2026-09-16T06:10:26.426305Z",
          "score": 100,
          "model": "gpt-6-1-sol",
          "runner": "benchmark_agent",
          "config": "openai-gpt-6-1-sol-xhigh-provider-adapter",
          "verified_publish": true,
          "levels_completed": 9,
          "actions": 547,
          "resets": 10,
          "state": "WIN",
          "completed": true,
          "number_of_levels": 0,
          "level_scores": [
            115,
            104.30058850158441,
            115,
            115,
            100,
            47.67573696145125,
            112.72671848803537,
            115,
            115
          ],
          "level_actions": [
            18,
            47,
            34,
            23,
            33,
            126,
            81,
            65,
            120
          ],
          "level_baseline_actions": [
            21,
            48,
            44,
            38,
            33,
            87,
            86,
            131,
            163
          ]
        },
        {
          "session_id": "ce3ffa36-65f1-4473-8630-3c4631a5f12a",
          "game_id": "bp35",
          "game_version": "0a0ad940",
          "published_at": "2026-09-16T05:44:48.280059Z",
          "score": 0.6786703601108035,
          "model": "gpt-6-1-sol",
          "runner": "benchmark_agent",
          "config": "openai-gpt-6-1-sol-medium",
          "verified_publish": true,
          "levels_completed": 1,
          "actions": 278,
          "resets": 47,
          "state": "GAME_OVER",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            30.540166204986157,
            0,
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            38,
            240,
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            21,
            48,
            44,
            38,
            33,
            87,
            86,
            131,
            163
          ]
        },
        {
          "session_id": "9e7a4382-f1cb-460e-a88a-498070e6a923",
          "game_id": "bp35",
          "game_version": "0a0ad940",
          "published_at": "2026-09-16T05:40:39.821972Z",
          "score": 1.7013888888888888,
          "model": "gpt-6-1-sol",
          "runner": "benchmark_agent",
          "config": "openai-gpt-6-1-sol-max",
          "verified_publish": true,
          "levels_completed": 1,
          "actions": 264,
          "resets": 52,
          "state": "GAME_OVER",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            76.5625,
            0,
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            24,
            240,
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            21,
            48,
            44,
            38,
            33,
            87,
            86,
            131,
            163
          ]
        },
        {
          "session_id": "4ded7d15-f758-48c1-b7ce-90c8b411fc6e",
          "game_id": "bp35",
          "game_version": "0a0ad940",
          "published_at": "2026-09-16T03:54:52.295523Z",
          "score": 0,
          "model": "gpt-6-1-sol",
          "runner": "benchmark_agent",
          "config": "openai-gpt-6-1-sol-high",
          "verified_publish": true,
          "levels_completed": 0,
          "actions": 105,
          "resets": 1,
          "state": "NOT_FINISHED",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            0,
            0,
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            105,
            0,
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            21,
            48,
            44,
            38,
            33,
            87,
            86,
            131,
            163
          ]
        },
        {
          "session_id": "083867c8-9854-41ec-839c-9db5d6729591",
          "game_id": "bp35",
          "game_version": "0a0ad940",
          "published_at": "2026-09-16T03:23:25.475907Z",
          "score": 1.4497041420118346,
          "model": "gpt-6-1-sol",
          "runner": "benchmark_agent",
          "config": "openai-gpt-6-1-sol-xhigh",
          "verified_publish": true,
          "levels_completed": 1,
          "actions": 266,
          "resets": 51,
          "state": "GAME_OVER",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            65.23668639053255,
            0,
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            26,
            240,
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            21,
            48,
            44,
            38,
            33,
            87,
            86,
            131,
            163
          ]
        },
        {
          "session_id": "14d3c1a3-e50b-4fb0-902a-518a8bbe601e",
          "game_id": "bp35",
          "game_version": "0a0ad940",
          "published_at": "2026-09-16T01:16:39.47456Z",
          "score": 100,
          "model": "gpt-6-1-sol",
          "runner": "benchmark_agent",
          "config": "openai-gpt-6-1-sol-max-provider-adapter",
          "verified_publish": true,
          "levels_completed": 9,
          "actions": 549,
          "resets": 7,
          "state": "WIN",
          "completed": true,
          "number_of_levels": 0,
          "level_scores": [
            115,
            42.07450693937181,
            115,
            115,
            113.3194588969823,
            38.61734693877551,
            115,
            115,
            115
          ],
          "level_actions": [
            16,
            74,
            34,
            23,
            31,
            140,
            54,
            53,
            124
          ],
          "level_baseline_actions": [
            21,
            48,
            44,
            38,
            33,
            87,
            86,
            131,
            163
          ]
        },
        {
          "session_id": "4930f4a6-e065-4d21-a133-790e4681e73b",
          "game_id": "bp35",
          "game_version": "0a0ad940",
          "published_at": "2026-09-11T05:16:31.283531Z",
          "score": 0,
          "model": "gemini-3.8-flash",
          "runner": "benchmark_agent",
          "config": "google-gemini-3-8-flash-low",
          "verified_publish": true,
          "levels_completed": 0,
          "actions": 105,
          "resets": 1,
          "state": "GAME_OVER",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            0,
            0,
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            105,
            0,
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            21,
            48,
            44,
            38,
            33,
            87,
            86,
            131,
            163
          ]
        },
        {
          "session_id": "856ca868-8d27-44be-b9e4-f4e8eec63cd9",
          "game_id": "bp35",
          "game_version": "0a0ad940",
          "published_at": "2026-09-10T16:56:49.558428Z",
          "score": 0,
          "model": "gemini-3.8-flash",
          "runner": "benchmark_agent",
          "config": "google-gemini-3-8-flash-medium",
          "verified_publish": true,
          "levels_completed": 0,
          "actions": 105,
          "resets": 1,
          "state": "GAME_OVER",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            0,
            0,
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            105,
            0,
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            21,
            48,
            44,
            38,
            33,
            87,
            86,
            131,
            163
          ]
        },
        {
          "session_id": "f3a75d2a-a38c-46a4-adb9-401a18ef5fc2",
          "game_id": "bp35",
          "game_version": "0a0ad940",
          "published_at": "2026-09-09T18:27:17.565319Z",
          "score": 0,
          "model": "gemini-3.8-flash",
          "runner": "benchmark_agent",
          "config": "google-gemini-3-8-flash-high",
          "verified_publish": true,
          "levels_completed": 0,
          "actions": 105,
          "resets": 1,
          "state": "GAME_OVER",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            0,
            0,
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            105,
            0,
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            21,
            48,
            44,
            38,
            33,
            87,
            86,
            131,
            163
          ]
        },
        {
          "session_id": "086aebf0-62f2-4f56-a809-87f0576036c0",
          "game_id": "bp35",
          "game_version": "0a0ad940",
          "published_at": "2026-08-11T07:38:57.097333Z",
          "score": 0,
          "model": "grok-4.6",
          "runner": "benchmark_agent",
          "config": "xai-grok-4-6-xhigh",
          "verified_publish": true,
          "levels_completed": 0,
          "actions": 105,
          "resets": 3,
          "state": "GAME_OVER",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            0,
            0,
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            105,
            0,
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            21,
            48,
            44,
            38,
            33,
            87,
            86,
            131,
            163
          ]
        },
        {
          "session_id": "ea232c70-c129-495d-b3e7-da1759333aeb",
          "game_id": "bp35",
          "game_version": "0a0ad940",
          "published_at": "2026-07-21T03:39:31.544642Z",
          "score": 0,
          "model": "claude-opus-5",
          "runner": "benchmark_agent",
          "config": "anthropic-claude-opus-5-high",
          "verified_publish": true,
          "levels_completed": 0,
          "actions": 105,
          "resets": 1,
          "state": "GAME_OVER",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            0,
            0,
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            105,
            0,
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            21,
            48,
            44,
            38,
            33,
            87,
            86,
            131,
            163
          ]
        },
        {
          "session_id": "de8d5543-bc54-4a40-b0d0-42ce5f4d009a",
          "game_id": "bp35",
          "game_version": "0a0ad940",
          "published_at": "2026-06-24T10:55:18.38066Z",
          "score": 0,
          "model": "gpt-5-6-luna",
          "runner": "benchmark_agent",
          "config": "openai-gpt-5-6-luna-low",
          "verified_publish": true,
          "levels_completed": 0,
          "actions": 105,
          "resets": 1,
          "state": "GAME_OVER",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            0,
            0,
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            105,
            0,
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            21,
            48,
            44,
            38,
            33,
            87,
            86,
            131,
            163
          ]
        },
        {
          "session_id": "55642d7d-967e-4257-a949-d000d5c58afb",
          "game_id": "bp35",
          "game_version": "0a0ad940",
          "published_at": "2026-06-24T10:34:16.683435Z",
          "score": 0,
          "model": "gpt-5-6-luna",
          "runner": "benchmark_agent",
          "config": "openai-gpt-5-6-luna-medium",
          "verified_publish": true,
          "levels_completed": 0,
          "actions": 105,
          "resets": 4,
          "state": "GAME_OVER",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            0,
            0,
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            105,
            0,
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            21,
            48,
            44,
            38,
            33,
            87,
            86,
            131,
            163
          ]
        },
        {
          "session_id": "b386bf3e-2c47-42c9-ad2c-7011abde7c66",
          "game_id": "bp35",
          "game_version": "0a0ad940",
          "published_at": "2026-06-24T03:43:13.247774Z",
          "score": 0,
          "model": "gpt-5-6-luna",
          "runner": "benchmark_agent",
          "config": "openai-gpt-5-6-luna-high",
          "verified_publish": true,
          "levels_completed": 0,
          "actions": 105,
          "resets": 5,
          "state": "GAME_OVER",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            0,
            0,
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            105,
            0,
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            21,
            48,
            44,
            38,
            33,
            87,
            86,
            131,
            163
          ]
        },
        {
          "session_id": "b3af0c80-09e8-445f-85df-b60bc558b629",
          "game_id": "bp35",
          "game_version": "0a0ad940",
          "published_at": "2026-06-24T01:31:33.889246Z",
          "score": 0,
          "model": "gpt-5-6-luna",
          "runner": "benchmark_agent",
          "config": "openai-gpt-5-6-luna-xhigh",
          "verified_publish": true,
          "levels_completed": 0,
          "actions": 105,
          "resets": 7,
          "state": "GAME_OVER",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            0,
            0,
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            105,
            0,
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            21,
            48,
            44,
            38,
            33,
            87,
            86,
            131,
            163
          ]
        },
        {
          "session_id": "4de6bc56-1208-4c87-b342-46e3054f64c6",
          "game_id": "bp35",
          "game_version": "0a0ad940",
          "published_at": "2026-06-23T01:34:44.284801Z",
          "score": 0,
          "model": "gpt-5-6-luna",
          "runner": "benchmark_agent",
          "config": "openai-gpt-5-6-luna-max",
          "verified_publish": true,
          "levels_completed": 0,
          "actions": 105,
          "resets": 9,
          "state": "GAME_OVER",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            0,
            0,
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            105,
            0,
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            21,
            48,
            44,
            38,
            33,
            87,
            86,
            131,
            163
          ]
        },
        {
          "session_id": "f5e8b272-f92c-4ba7-b639-b80bb300c4ca",
          "game_id": "bp35",
          "game_version": "0a0ad940",
          "published_at": "2026-06-20T12:01:15.968331Z",
          "score": 0,
          "model": "gpt-5-6-terra",
          "runner": "benchmark_agent",
          "config": "openai-gpt-5-6-terra-low",
          "verified_publish": true,
          "levels_completed": 0,
          "actions": 105,
          "resets": 2,
          "state": "GAME_OVER",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            0,
            0,
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            105,
            0,
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            21,
            48,
            44,
            38,
            33,
            87,
            86,
            131,
            163
          ]
        },
        {
          "session_id": "04a8ae17-22b5-47ab-b74e-2e5ec7425b29",
          "game_id": "bp35",
          "game_version": "0a0ad940",
          "published_at": "2026-06-20T09:22:58.590048Z",
          "score": 0,
          "model": "gpt-5-6-terra",
          "runner": "benchmark_agent",
          "config": "openai-gpt-5-6-terra-medium",
          "verified_publish": true,
          "levels_completed": 0,
          "actions": 105,
          "resets": 1,
          "state": "GAME_OVER",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            0,
            0,
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            105,
            0,
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            21,
            48,
            44,
            38,
            33,
            87,
            86,
            131,
            163
          ]
        },
        {
          "session_id": "5b1a9f1a-1ba7-4ead-a71d-fb5fc3176faa",
          "game_id": "bp35",
          "game_version": "0a0ad940",
          "published_at": "2026-06-20T08:42:04.913229Z",
          "score": 0,
          "model": "gpt-5-6-terra",
          "runner": "benchmark_agent",
          "config": "openai-gpt-5-6-terra-high",
          "verified_publish": true,
          "levels_completed": 0,
          "actions": 105,
          "resets": 1,
          "state": "GAME_OVER",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            0,
            0,
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            105,
            0,
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            21,
            48,
            44,
            38,
            33,
            87,
            86,
            131,
            163
          ]
        },
        {
          "session_id": "91dbe690-a73a-4a4c-ad58-686ca74b78f1",
          "game_id": "bp35",
          "game_version": "0a0ad940",
          "published_at": "2026-06-20T02:08:41.936779Z",
          "score": 0,
          "model": "gpt-5-6-terra",
          "runner": "benchmark_agent",
          "config": "openai-gpt-5-6-terra-xhigh",
          "verified_publish": true,
          "levels_completed": 0,
          "actions": 105,
          "resets": 1,
          "state": "GAME_OVER",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            0,
            0,
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            105,
            0,
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            21,
            48,
            44,
            38,
            33,
            87,
            86,
            131,
            163
          ]
        },
        {
          "session_id": "db096fe5-3caa-4cd2-b64e-bb4aa9a67cd5",
          "game_id": "bp35",
          "game_version": "0a0ad940",
          "published_at": "2026-06-18T21:25:47.671407Z",
          "score": 0,
          "model": "gpt-5-6-terra",
          "runner": "benchmark_agent",
          "config": "openai-gpt-5-6-terra-max",
          "verified_publish": true,
          "levels_completed": 0,
          "actions": 105,
          "resets": 1,
          "state": "GAME_OVER",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            0,
            0,
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            105,
            0,
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            21,
            48,
            44,
            38,
            33,
            87,
            86,
            131,
            163
          ]
        },
        {
          "session_id": "5a2a4b50-bb1b-4219-a268-35dee2268d28",
          "game_id": "bp35",
          "game_version": "0a0ad940",
          "published_at": "2026-06-13T02:35:19.027526Z",
          "score": 0,
          "model": "gpt-5-6-sol",
          "runner": "benchmark_agent",
          "config": "openai-gpt-5-6-sol-max",
          "verified_publish": true,
          "levels_completed": 0,
          "actions": 105,
          "resets": 1,
          "state": "GAME_OVER",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            0,
            0,
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            105,
            0,
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            21,
            48,
            44,
            38,
            33,
            87,
            86,
            131,
            163
          ]
        },
        {
          "session_id": "41c5b3f6-5667-4c6d-9aeb-314ab96d56a1",
          "game_id": "bp35",
          "game_version": "0a0ad940",
          "published_at": "2026-06-11T16:11:22.523226Z",
          "score": 0,
          "model": "gpt-5-6-sol",
          "runner": "benchmark_agent",
          "config": "openai-gpt-5-6-sol-xhigh",
          "verified_publish": true,
          "levels_completed": 0,
          "actions": 105,
          "resets": 1,
          "state": "GAME_OVER",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            0,
            0,
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            105,
            0,
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            21,
            48,
            44,
            38,
            33,
            87,
            86,
            131,
            163
          ]
        },
        {
          "session_id": "5ad88a45-3663-4a62-8af9-a133bc32def1",
          "game_id": "bp35",
          "game_version": "0a0ad940",
          "published_at": "2026-06-10T11:47:12.83387Z",
          "score": 0,
          "model": "gpt-5-6-sol",
          "runner": "benchmark_agent",
          "config": "openai-gpt-5-6-sol-high",
          "verified_publish": true,
          "levels_completed": 0,
          "actions": 105,
          "resets": 2,
          "state": "GAME_OVER",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            0,
            0,
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            105,
            0,
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            21,
            48,
            44,
            38,
            33,
            87,
            86,
            131,
            163
          ]
        },
        {
          "session_id": "26c2bf2e-acfc-4be2-a7f8-feab586eed36",
          "game_id": "bp35",
          "game_version": "0a0ad940",
          "published_at": "2026-06-10T08:35:37.14512Z",
          "score": 0,
          "model": "gpt-5-6-sol",
          "runner": "benchmark_agent",
          "config": "openai-gpt-5-6-sol-medium",
          "verified_publish": true,
          "levels_completed": 0,
          "actions": 105,
          "resets": 2,
          "state": "GAME_OVER",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            0,
            0,
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            105,
            0,
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            21,
            48,
            44,
            38,
            33,
            87,
            86,
            131,
            163
          ]
        },
        {
          "session_id": "c6fc13b8-ccb3-44ca-b33f-e2058901f16f",
          "game_id": "bp35",
          "game_version": "0a0ad940",
          "published_at": "2026-06-09T22:12:32.505958Z",
          "score": 0,
          "model": "gpt-5-6-sol",
          "runner": "benchmark_agent",
          "config": "openai-gpt-5-6-sol-low",
          "verified_publish": true,
          "levels_completed": 0,
          "actions": 105,
          "resets": 2,
          "state": "GAME_OVER",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            0,
            0,
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            105,
            0,
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            21,
            48,
            44,
            38,
            33,
            87,
            86,
            131,
            163
          ]
        },
        {
          "session_id": "be02a5af-f904-45d4-8c6d-ca991f0f6994",
          "game_id": "bp35",
          "game_version": "0a0ad940",
          "published_at": "2026-05-29T17:13:35.657066Z",
          "score": 0,
          "model": "claude-opus-4-8",
          "runner": "benchmark_agent",
          "config": "anthropic-opus-4-8-high",
          "verified_publish": true,
          "levels_completed": 0,
          "actions": 105,
          "resets": 1,
          "state": "NOT_FINISHED",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            0,
            0,
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            105,
            0,
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            21,
            48,
            44,
            38,
            33,
            87,
            86,
            131,
            163
          ]
        },
        {
          "session_id": "125f6e8b-4258-431a-80c1-9bd945f60dca",
          "game_id": "bp35",
          "game_version": "0a0ad940",
          "published_at": "2026-05-01T16:14:34.368836Z",
          "score": 0,
          "model": "gpt-5.5-2026-04-23",
          "runner": "benchmark_agent",
          "config": "openai-gpt-5-5-2026-04-23-high",
          "verified_publish": true,
          "levels_completed": 0,
          "actions": 105,
          "resets": 2,
          "state": "GAME_OVER",
          "completed": false,
          "number_of_levels": 9,
          "level_scores": [
            0,
            0,
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            105,
            0,
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            21,
            48,
            44,
            38,
            33,
            87,
            86,
            131,
            163
          ]
        },
        {
          "session_id": "defddeee-6257-4349-b2d7-7ace0ee08d07",
          "game_id": "bp35",
          "game_version": "0a0ad940",
          "published_at": "2026-05-01T05:07:14.140778Z",
          "score": 0,
          "model": "claude-opus-4-7",
          "runner": "benchmark_agent",
          "config": "anthropic-opus-4-7-high",
          "verified_publish": true,
          "levels_completed": 0,
          "actions": 105,
          "resets": 1,
          "state": "GAME_OVER",
          "completed": false,
          "number_of_levels": 9,
          "level_scores": [
            0,
            0,
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            105,
            0,
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            21,
            48,
            44,
            38,
            33,
            87,
            86,
            131,
            163
          ]
        },
        {
          "session_id": "a64ac589-a3fa-410d-8afe-0fbe6aa34ea3",
          "game_id": "bp35",
          "game_version": "0a0ad940",
          "published_at": "2026-04-14T17:44:10.808836Z",
          "score": 0,
          "model": "grok-4.20-beta-0309-reasoning",
          "runner": "benchmark_agent",
          "config": "xai-grok-4-20-beta-0309-reasoning",
          "verified_publish": true,
          "levels_completed": 0,
          "actions": 105,
          "resets": 1,
          "state": "GAME_OVER",
          "completed": false,
          "number_of_levels": 9,
          "level_scores": [
            0,
            0,
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            105,
            0,
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            21,
            48,
            44,
            38,
            33,
            87,
            86,
            131,
            163
          ]
        },
        {
          "session_id": "cb8babe3-db01-4d89-945f-21548ac566b5",
          "game_id": "bp35",
          "game_version": "0a0ad940",
          "published_at": "2026-04-14T17:44:10.44061Z",
          "score": 0,
          "model": "gpt-5.4-2026-03-05",
          "runner": "benchmark_agent",
          "config": "openai-gpt-5-4-2026-03-05-high",
          "verified_publish": true,
          "levels_completed": 0,
          "actions": 105,
          "resets": 2,
          "state": "GAME_OVER",
          "completed": false,
          "number_of_levels": 9,
          "level_scores": [
            0,
            0,
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            105,
            0,
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            21,
            48,
            44,
            38,
            33,
            87,
            86,
            131,
            163
          ]
        },
        {
          "session_id": "44bc1e40-9757-43d9-938e-4b217e51274d",
          "game_id": "bp35",
          "game_version": "0a0ad940",
          "published_at": "2026-04-14T17:44:09.636826Z",
          "score": 0,
          "model": "gemini-3.1-pro-preview",
          "runner": "benchmark_agent",
          "config": "google-gemini-3-1-pro-preview",
          "verified_publish": true,
          "levels_completed": 0,
          "actions": 105,
          "resets": 2,
          "state": "GAME_OVER",
          "completed": false,
          "number_of_levels": 9,
          "level_scores": [
            0,
            0,
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            105,
            0,
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            21,
            48,
            44,
            38,
            33,
            87,
            86,
            131,
            163
          ]
        },
        {
          "session_id": "2317d132-4f08-4076-aca2-fde14c949001",
          "game_id": "bp35",
          "game_version": "0a0ad940",
          "published_at": "2026-04-14T17:44:09.236576Z",
          "score": 0,
          "model": "claude-opus-4-6",
          "runner": "benchmark_agent",
          "config": "anthropic-opus-4-6-max-effort",
          "verified_publish": true,
          "levels_completed": 0,
          "actions": 105,
          "resets": 3,
          "state": "GAME_OVER",
          "completed": false,
          "number_of_levels": 9,
          "level_scores": [
            0,
            0,
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            105,
            0,
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            21,
            48,
            44,
            38,
            33,
            87,
            86,
            131,
            163
          ]
        }
      ]
    },
    {
      "game": "dc22",
      "source": "https://arcprize.org/tasks/dc22",
      "runs": [
        {
          "session_id": "751f2853-48e0-4140-811c-9f614d48c167",
          "game_id": "dc22",
          "game_version": "fdcac232",
          "published_at": "2026-09-25T15:27:06.607941Z",
          "score": 22.956080372710787,
          "model": "gpt-6-sol",
          "runner": "benchmark_agent",
          "config": "openai-gpt-6-sol-max",
          "verified_publish": true,
          "levels_completed": 3,
          "actions": 697,
          "resets": 2,
          "state": "GAME_OVER",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            115,
            11.038843913463271,
            115,
            0,
            0,
            0
          ],
          "level_actions": [
            27,
            307,
            59,
            304,
            0,
            0
          ],
          "level_baseline_actions": [
            59,
            102,
            67,
            98,
            324,
            578
          ]
        },
        {
          "session_id": "7774b2ea-48ff-4bc8-9fae-293dd54ac60c",
          "game_id": "dc22",
          "game_version": "fdcac232",
          "published_at": "2026-09-25T13:12:23.121826Z",
          "score": 0,
          "model": "gpt-6-sol",
          "runner": "benchmark_agent",
          "config": "openai-gpt-6-sol-none",
          "verified_publish": true,
          "levels_completed": 0,
          "actions": 295,
          "resets": 2,
          "state": "GAME_OVER",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            295,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            59,
            102,
            67,
            98,
            324,
            578
          ]
        },
        {
          "session_id": "b86c67af-317f-4f86-bc36-d98ba52fd346",
          "game_id": "dc22",
          "game_version": "fdcac232",
          "published_at": "2026-09-25T12:48:25.892564Z",
          "score": 4.86536627274324,
          "model": "gpt-6-sol",
          "runner": "benchmark_agent",
          "config": "openai-gpt-6-sol-none-provider-adapter",
          "verified_publish": true,
          "levels_completed": 2,
          "actions": 687,
          "resets": 7,
          "state": "GAME_OVER",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            8.447183867601735,
            46.86275393000315,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            203,
            149,
            335,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            59,
            102,
            67,
            98,
            324,
            578
          ]
        },
        {
          "session_id": "829d628d-9008-46df-9f4f-f591d5f02429",
          "game_id": "dc22",
          "game_version": "fdcac232",
          "published_at": "2026-09-25T07:06:36.492851Z",
          "score": 17.87547666740855,
          "model": "gpt-6-sol",
          "runner": "benchmark_agent",
          "config": "openai-gpt-6-sol-medium-provider-adapter",
          "verified_publish": true,
          "levels_completed": 3,
          "actions": 779,
          "resets": 11,
          "state": "NOT_FINISHED",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            82.39053254437869,
            101.99000098029605,
            29.671491836869585,
            0,
            0,
            0
          ],
          "level_actions": [
            65,
            101,
            123,
            490,
            0,
            0
          ],
          "level_baseline_actions": [
            59,
            102,
            67,
            98,
            324,
            578
          ]
        },
        {
          "session_id": "e9f9be7d-7a1f-49c2-8d6f-f3595762cf48",
          "game_id": "dc22",
          "game_version": "fdcac232",
          "published_at": "2026-09-25T07:04:02.803083Z",
          "score": 0.37232295941669047,
          "model": "gpt-6-sol",
          "runner": "benchmark_agent",
          "config": "openai-gpt-6-sol-low-provider-adapter",
          "verified_publish": true,
          "levels_completed": 1,
          "actions": 721,
          "resets": 9,
          "state": "GAME_OVER",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            7.8187821477504995,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            211,
            510,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            59,
            102,
            67,
            98,
            324,
            578
          ]
        },
        {
          "session_id": "46cec91c-7eaf-4261-a165-9ca92180580b",
          "game_id": "dc22",
          "game_version": "fdcac232",
          "published_at": "2026-09-25T06:42:05.265388Z",
          "score": 0,
          "model": "gpt-6-sol",
          "runner": "benchmark_agent",
          "config": "openai-gpt-6-sol-xhigh",
          "verified_publish": true,
          "levels_completed": 0,
          "actions": 295,
          "resets": 2,
          "state": "GAME_OVER",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            295,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            59,
            102,
            67,
            98,
            324,
            578
          ]
        },
        {
          "session_id": "086b575c-d1b0-4306-9306-c5c12d5ee8f3",
          "game_id": "dc22",
          "game_version": "fdcac232",
          "published_at": "2026-09-25T06:24:45.900391Z",
          "score": 16.16238900571388,
          "model": "gpt-6-sol",
          "runner": "benchmark_agent",
          "config": "openai-gpt-6-sol-high-provider-adapter",
          "verified_publish": true,
          "levels_completed": 3,
          "actions": 782,
          "resets": 9,
          "state": "NOT_FINISHED",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            115,
            40.640625,
            47.70963970666383,
            0,
            0,
            0
          ],
          "level_actions": [
            35,
            160,
            97,
            490,
            0,
            0
          ],
          "level_baseline_actions": [
            59,
            102,
            67,
            98,
            324,
            578
          ]
        },
        {
          "session_id": "995654f0-0b43-470a-bafc-7e60ad7921f8",
          "game_id": "dc22",
          "game_version": "fdcac232",
          "published_at": "2026-09-25T06:20:08.173762Z",
          "score": 0,
          "model": "gpt-6-sol",
          "runner": "benchmark_agent",
          "config": "openai-gpt-6-sol-low",
          "verified_publish": true,
          "levels_completed": 0,
          "actions": 295,
          "resets": 2,
          "state": "GAME_OVER",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            295,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            59,
            102,
            67,
            98,
            324,
            578
          ]
        },
        {
          "session_id": "d7d016f4-123e-4e7d-ad7b-04b9b771ddee",
          "game_id": "dc22",
          "game_version": "fdcac232",
          "published_at": "2026-09-25T06:06:32.690965Z",
          "score": 17.620020808561236,
          "model": "gpt-6-sol",
          "runner": "benchmark_agent",
          "config": "openai-gpt-6-sol-xhigh-provider-adapter",
          "verified_publish": true,
          "levels_completed": 3,
          "actions": 824,
          "resets": 8,
          "state": "GAME_OVER",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            115,
            115,
            8.340145659928657,
            0,
            0,
            0
          ],
          "level_actions": [
            29,
            73,
            232,
            490,
            0,
            0
          ],
          "level_baseline_actions": [
            59,
            102,
            67,
            98,
            324,
            578
          ]
        },
        {
          "session_id": "3b973d0c-902c-43b5-9c86-64a75671dbc2",
          "game_id": "dc22",
          "game_version": "fdcac232",
          "published_at": "2026-09-25T06:06:21.587769Z",
          "score": 0,
          "model": "gpt-6-sol",
          "runner": "benchmark_agent",
          "config": "openai-gpt-6-sol-medium",
          "verified_publish": true,
          "levels_completed": 0,
          "actions": 295,
          "resets": 2,
          "state": "GAME_OVER",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            295,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            59,
            102,
            67,
            98,
            324,
            578
          ]
        },
        {
          "session_id": "b9f1be70-4c02-4c50-a46a-6ca94bef3fcc",
          "game_id": "dc22",
          "game_version": "fdcac232",
          "published_at": "2026-09-25T05:35:53.339167Z",
          "score": 0,
          "model": "gpt-6-sol",
          "runner": "benchmark_agent",
          "config": "openai-gpt-6-sol-high",
          "verified_publish": true,
          "levels_completed": 0,
          "actions": 295,
          "resets": 3,
          "state": "GAME_OVER",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            295,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            59,
            102,
            67,
            98,
            324,
            578
          ]
        },
        {
          "session_id": "333efe22-785f-4474-9581-9e29ad87f5e7",
          "game_id": "dc22",
          "game_version": "fdcac232",
          "published_at": "2026-09-21T11:58:44.572434Z",
          "score": 45.76677248677249,
          "model": "gemini-3.8-flash",
          "runner": "benchmark_agent",
          "config": "google-gemini-3-8-flash-low-provider-adapter",
          "verified_publish": true,
          "levels_completed": 4,
          "actions": 2026,
          "resets": 6,
          "state": "GAME_OVER",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            115,
            20.55111111111111,
            115,
            115,
            0,
            0
          ],
          "level_actions": [
            48,
            225,
            58,
            75,
            1620,
            0
          ],
          "level_baseline_actions": [
            59,
            102,
            67,
            98,
            324,
            578
          ]
        },
        {
          "session_id": "05c13d45-7584-47f8-8831-42d99baef060",
          "game_id": "dc22",
          "game_version": "fdcac232",
          "published_at": "2026-09-20T07:50:16.599613Z",
          "score": 47.61904761904761,
          "model": "gemini-3.8-flash",
          "runner": "benchmark_agent",
          "config": "google-gemini-3-8-flash-medium-provider-adapter",
          "verified_publish": true,
          "levels_completed": 4,
          "actions": 2010,
          "resets": 5,
          "state": "GAME_OVER",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            8.790182066109441,
            115,
            106.24852071005915,
            115,
            0,
            0
          ],
          "level_actions": [
            199,
            57,
            65,
            69,
            1620,
            0
          ],
          "level_baseline_actions": [
            59,
            102,
            67,
            98,
            324,
            578
          ]
        },
        {
          "session_id": "dca57f83-baae-4af5-ba83-b430768c6589",
          "game_id": "dc22",
          "game_version": "fdcac232",
          "published_at": "2026-09-18T02:08:31.906851Z",
          "score": 54.59350981783328,
          "model": "gemini-3.8-flash",
          "runner": "benchmark_agent",
          "config": "google-gemini-3-8-flash-high-provider-adapter",
          "verified_publish": true,
          "levels_completed": 5,
          "actions": 1820,
          "resets": 5,
          "state": "GAME_OVER",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            115,
            115,
            106.24852071005915,
            115,
            4.543628808864266,
            0
          ],
          "level_actions": [
            24,
            45,
            65,
            71,
            1520,
            95
          ],
          "level_baseline_actions": [
            59,
            102,
            67,
            98,
            324,
            578
          ]
        },
        {
          "session_id": "d7ce15d8-19c7-41cd-9f8c-d5a1303664da",
          "game_id": "dc22",
          "game_version": "fdcac232",
          "published_at": "2026-09-16T20:49:04.783391Z",
          "score": 47.61904761904761,
          "model": "gpt-6-1-sol",
          "runner": "benchmark_agent",
          "config": "openai-gpt-6-1-sol-high",
          "verified_publish": true,
          "levels_completed": 4,
          "actions": 1844,
          "resets": 5,
          "state": "GAME_OVER",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            115,
            115,
            115,
            115,
            0,
            0
          ],
          "level_actions": [
            24,
            75,
            55,
            70,
            1620,
            0
          ],
          "level_baseline_actions": [
            59,
            102,
            67,
            98,
            324,
            578
          ]
        },
        {
          "session_id": "4c7b5078-f640-4fe6-85f8-53413cce3336",
          "game_id": "dc22",
          "game_version": "fdcac232",
          "published_at": "2026-09-16T20:02:29.910712Z",
          "score": 0,
          "model": "gpt-6-luna",
          "runner": "benchmark_agent",
          "config": "openai-gpt-6-luna-medium",
          "verified_publish": true,
          "levels_completed": 0,
          "actions": 295,
          "resets": 4,
          "state": "GAME_OVER",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            295,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            59,
            102,
            67,
            98,
            324,
            578
          ]
        },
        {
          "session_id": "963d205b-f30e-48ed-ad12-fe3fb078e11b",
          "game_id": "dc22",
          "game_version": "fdcac232",
          "published_at": "2026-09-16T19:39:01.807199Z",
          "score": 0,
          "model": "gpt-6-luna",
          "runner": "benchmark_agent",
          "config": "openai-gpt-6-luna-low",
          "verified_publish": true,
          "levels_completed": 0,
          "actions": 295,
          "resets": 9,
          "state": "GAME_OVER",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            295,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            59,
            102,
            67,
            98,
            324,
            578
          ]
        },
        {
          "session_id": "7170fc3a-2b78-4e03-933a-8831d87cccd5",
          "game_id": "dc22",
          "game_version": "fdcac232",
          "published_at": "2026-09-16T19:34:29.152137Z",
          "score": 0,
          "model": "gpt-6-luna",
          "runner": "benchmark_agent",
          "config": "openai-gpt-6-luna-low-provider-adapter",
          "verified_publish": true,
          "levels_completed": 0,
          "actions": 295,
          "resets": 3,
          "state": "GAME_OVER",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            295,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            59,
            102,
            67,
            98,
            324,
            578
          ]
        },
        {
          "session_id": "3cc37b99-a7bb-4dfe-86b0-bf3c46e69829",
          "game_id": "dc22",
          "game_version": "fdcac232",
          "published_at": "2026-09-16T19:30:14.514974Z",
          "score": 0,
          "model": "gpt-6-luna",
          "runner": "benchmark_agent",
          "config": "openai-gpt-6-luna-none-provider-adapter",
          "verified_publish": true,
          "levels_completed": 0,
          "actions": 295,
          "resets": 9,
          "state": "NOT_FINISHED",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            295,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            59,
            102,
            67,
            98,
            324,
            578
          ]
        },
        {
          "session_id": "16ba74a0-51b7-4640-8063-3504e3b07b6a",
          "game_id": "dc22",
          "game_version": "fdcac232",
          "published_at": "2026-09-16T19:29:17.866303Z",
          "score": 0,
          "model": "gpt-6-luna",
          "runner": "benchmark_agent",
          "config": "openai-gpt-6-luna-none",
          "verified_publish": true,
          "levels_completed": 0,
          "actions": 295,
          "resets": 14,
          "state": "NOT_FINISHED",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            295,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            59,
            102,
            67,
            98,
            324,
            578
          ]
        },
        {
          "session_id": "c251dd88-1444-4fe2-8477-c060de616d8e",
          "game_id": "dc22",
          "game_version": "fdcac232",
          "published_at": "2026-09-16T16:54:38.23755Z",
          "score": 0,
          "model": "gpt-6-luna",
          "runner": "benchmark_agent",
          "config": "openai-gpt-6-luna-xhigh",
          "verified_publish": true,
          "levels_completed": 0,
          "actions": 295,
          "resets": 3,
          "state": "GAME_OVER",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            295,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            59,
            102,
            67,
            98,
            324,
            578
          ]
        },
        {
          "session_id": "46baf0f4-f738-4785-9025-87f3b9962291",
          "game_id": "dc22",
          "game_version": "fdcac232",
          "published_at": "2026-09-16T16:15:37.236993Z",
          "score": 0,
          "model": "gpt-6-luna",
          "runner": "benchmark_agent",
          "config": "openai-gpt-6-luna-high-provider-adapter",
          "verified_publish": true,
          "levels_completed": 0,
          "actions": 295,
          "resets": 2,
          "state": "GAME_OVER",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            295,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            59,
            102,
            67,
            98,
            324,
            578
          ]
        },
        {
          "session_id": "76ec2178-9770-4b51-aa03-bd3a58f45020",
          "game_id": "dc22",
          "game_version": "fdcac232",
          "published_at": "2026-09-16T16:07:33.377366Z",
          "score": 0,
          "model": "gpt-6-luna",
          "runner": "benchmark_agent",
          "config": "openai-gpt-6-luna-high",
          "verified_publish": true,
          "levels_completed": 0,
          "actions": 295,
          "resets": 4,
          "state": "GAME_OVER",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            295,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            59,
            102,
            67,
            98,
            324,
            578
          ]
        },
        {
          "session_id": "4d4c5751-90a1-434d-b4e1-5402d24fdb67",
          "game_id": "dc22",
          "game_version": "fdcac232",
          "published_at": "2026-09-16T09:33:34.496046Z",
          "score": 0,
          "model": "gpt-6-luna",
          "runner": "benchmark_agent",
          "config": "openai-gpt-6-luna-medium-provider-adapter",
          "verified_publish": true,
          "levels_completed": 0,
          "actions": 295,
          "resets": 2,
          "state": "GAME_OVER",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            295,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            59,
            102,
            67,
            98,
            324,
            578
          ]
        },
        {
          "session_id": "3d0d3c46-b0ff-4649-a4b3-d0c0f0bf3a4d",
          "game_id": "dc22",
          "game_version": "fdcac232",
          "published_at": "2026-09-16T09:11:50.162811Z",
          "score": 0,
          "model": "gpt-6-luna",
          "runner": "benchmark_agent",
          "config": "openai-gpt-6-luna-max-provider-adapter",
          "verified_publish": true,
          "levels_completed": 0,
          "actions": 295,
          "resets": 3,
          "state": "NOT_FINISHED",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            295,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            59,
            102,
            67,
            98,
            324,
            578
          ]
        },
        {
          "session_id": "b60cc853-a5ab-4d37-8908-c1ac46d898b9",
          "game_id": "dc22",
          "game_version": "fdcac232",
          "published_at": "2026-09-16T09:08:39.661137Z",
          "score": 47.61904761904761,
          "model": "gpt-6-1-sol",
          "runner": "benchmark_agent",
          "config": "openai-gpt-6-1-sol-xhigh",
          "verified_publish": true,
          "levels_completed": 4,
          "actions": 1328,
          "resets": 3,
          "state": "GAME_OVER",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            115,
            115,
            115,
            115,
            0,
            0
          ],
          "level_actions": [
            24,
            45,
            49,
            65,
            1145,
            0
          ],
          "level_baseline_actions": [
            59,
            102,
            67,
            98,
            324,
            578
          ]
        },
        {
          "session_id": "50b020b4-3bfb-4a79-854b-5251ba50da61",
          "game_id": "dc22",
          "game_version": "fdcac232",
          "published_at": "2026-09-16T08:41:08.391018Z",
          "score": 0,
          "model": "gpt-6-luna",
          "runner": "benchmark_agent",
          "config": "openai-gpt-6-luna-xhigh-provider-adapter",
          "verified_publish": true,
          "levels_completed": 0,
          "actions": 295,
          "resets": 3,
          "state": "GAME_OVER",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            295,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            59,
            102,
            67,
            98,
            324,
            578
          ]
        },
        {
          "session_id": "6fb2cbc2-8e6d-4e63-9256-59762bcf14a4",
          "game_id": "dc22",
          "game_version": "fdcac232",
          "published_at": "2026-09-16T08:36:28.111498Z",
          "score": 39.20653183864469,
          "model": "gpt-6-1-sol",
          "runner": "benchmark_agent",
          "config": "openai-gpt-6-1-sol-medium",
          "verified_publish": true,
          "levels_completed": 4,
          "actions": 2048,
          "resets": 6,
          "state": "GAME_OVER",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            115,
            115,
            6.112389537179505,
            115,
            0,
            0
          ],
          "level_actions": [
            26,
            64,
            271,
            67,
            1620,
            0
          ],
          "level_baseline_actions": [
            59,
            102,
            67,
            98,
            324,
            578
          ]
        },
        {
          "session_id": "39696b27-b27c-42d2-a173-e9a9126ab0d2",
          "game_id": "dc22",
          "game_version": "fdcac232",
          "published_at": "2026-09-16T07:40:55.990414Z",
          "score": 0,
          "model": "gpt-6-luna",
          "runner": "benchmark_agent",
          "config": "openai-gpt-6-luna-max",
          "verified_publish": true,
          "levels_completed": 0,
          "actions": 295,
          "resets": 4,
          "state": "GAME_OVER",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            295,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            59,
            102,
            67,
            98,
            324,
            578
          ]
        },
        {
          "session_id": "94b78163-27ee-426c-b697-b438851b06c5",
          "game_id": "dc22",
          "game_version": "fdcac232",
          "published_at": "2026-09-16T07:27:19.18714Z",
          "score": 47.61904761904761,
          "model": "gpt-6-1-sol",
          "runner": "benchmark_agent",
          "config": "openai-gpt-6-1-sol-low",
          "verified_publish": true,
          "levels_completed": 4,
          "actions": 445,
          "resets": 0,
          "state": "NOT_FINISHED",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            115,
            115,
            115,
            115,
            0,
            0
          ],
          "level_actions": [
            25,
            84,
            59,
            71,
            206,
            0
          ],
          "level_baseline_actions": [
            59,
            102,
            67,
            98,
            324,
            578
          ]
        },
        {
          "session_id": "d2f668aa-e056-43ee-8b08-ad32dbf57068",
          "game_id": "dc22",
          "game_version": "fdcac232",
          "published_at": "2026-09-16T05:46:19.804867Z",
          "score": 100,
          "model": "gpt-6-1-sol",
          "runner": "benchmark_agent",
          "config": "openai-gpt-6-1-sol-high-provider-adapter",
          "verified_publish": true,
          "levels_completed": 6,
          "actions": 497,
          "resets": 0,
          "state": "WIN",
          "completed": true,
          "number_of_levels": 0,
          "level_scores": [
            115,
            115,
            115,
            115,
            115,
            115
          ],
          "level_actions": [
            33,
            46,
            45,
            63,
            131,
            179
          ],
          "level_baseline_actions": [
            59,
            102,
            67,
            98,
            324,
            578
          ]
        },
        {
          "session_id": "c4c6125d-305d-4f9d-bea3-741d58b2831a",
          "game_id": "dc22",
          "game_version": "fdcac232",
          "published_at": "2026-09-16T04:23:02.193362Z",
          "score": 100,
          "model": "gpt-6-1-sol",
          "runner": "benchmark_agent",
          "config": "openai-gpt-6-1-sol-max-provider-adapter",
          "verified_publish": true,
          "levels_completed": 6,
          "actions": 459,
          "resets": 0,
          "state": "WIN",
          "completed": true,
          "number_of_levels": 0,
          "level_scores": [
            115,
            115,
            115,
            115,
            115,
            115
          ],
          "level_actions": [
            22,
            44,
            46,
            64,
            125,
            158
          ],
          "level_baseline_actions": [
            59,
            102,
            67,
            98,
            324,
            578
          ]
        },
        {
          "session_id": "73658101-90b5-4cc3-9bfa-a4124485cbcb",
          "game_id": "dc22",
          "game_version": "fdcac232",
          "published_at": "2026-09-16T03:35:29.285385Z",
          "score": 100,
          "model": "gpt-6-1-sol",
          "runner": "benchmark_agent",
          "config": "openai-gpt-6-1-sol-xhigh-provider-adapter",
          "verified_publish": true,
          "levels_completed": 6,
          "actions": 471,
          "resets": 0,
          "state": "WIN",
          "completed": true,
          "number_of_levels": 0,
          "level_scores": [
            115,
            115,
            115,
            115,
            115,
            115
          ],
          "level_actions": [
            30,
            42,
            52,
            72,
            117,
            158
          ],
          "level_baseline_actions": [
            59,
            102,
            67,
            98,
            324,
            578
          ]
        },
        {
          "session_id": "1f1db2e6-3158-42dd-985c-31c60f015962",
          "game_id": "dc22",
          "game_version": "fdcac232",
          "published_at": "2026-09-16T03:21:43.786144Z",
          "score": 100,
          "model": "gpt-6-1-sol",
          "runner": "benchmark_agent",
          "config": "openai-gpt-6-1-sol-max",
          "verified_publish": true,
          "levels_completed": 6,
          "actions": 751,
          "resets": 0,
          "state": "WIN",
          "completed": true,
          "number_of_levels": 0,
          "level_scores": [
            115,
            115,
            115,
            115,
            115,
            115
          ],
          "level_actions": [
            24,
            43,
            47,
            63,
            143,
            431
          ],
          "level_baseline_actions": [
            59,
            102,
            67,
            98,
            324,
            578
          ]
        },
        {
          "session_id": "22657662-52c6-448e-bcc7-f6b26d518524",
          "game_id": "dc22",
          "game_version": "fdcac232",
          "published_at": "2026-09-12T16:00:05.864148Z",
          "score": 47.61904761904761,
          "model": "gemini-3.8-flash",
          "runner": "benchmark_agent",
          "config": "google-gemini-3-8-flash-high",
          "verified_publish": true,
          "levels_completed": 4,
          "actions": 1324,
          "resets": 3,
          "state": "GAME_OVER",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            115,
            115,
            115,
            115,
            0,
            0
          ],
          "level_actions": [
            45,
            45,
            50,
            73,
            1111,
            0
          ],
          "level_baseline_actions": [
            59,
            102,
            67,
            98,
            324,
            578
          ]
        },
        {
          "session_id": "931ccf97-d68b-4c2e-98f8-e290aadd776d",
          "game_id": "dc22",
          "game_version": "fdcac232",
          "published_at": "2026-09-12T04:29:07.816329Z",
          "score": 23.97685630458539,
          "model": "gemini-3.8-flash",
          "runner": "benchmark_agent",
          "config": "google-gemini-3-8-flash-low",
          "verified_publish": true,
          "levels_completed": 3,
          "actions": 741,
          "resets": 4,
          "state": "GAME_OVER",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            111.0012755102041,
            69.90056436441817,
            84.23719271908425,
            0,
            0,
            0
          ],
          "level_actions": [
            56,
            122,
            73,
            490,
            0,
            0
          ],
          "level_baseline_actions": [
            59,
            102,
            67,
            98,
            324,
            578
          ]
        },
        {
          "session_id": "f2370245-8403-4658-b008-e06b01bd0287",
          "game_id": "dc22",
          "game_version": "fdcac232",
          "published_at": "2026-09-10T19:55:24.238262Z",
          "score": 4.761904761904762,
          "model": "gemini-3.8-flash",
          "runner": "benchmark_agent",
          "config": "google-gemini-3-8-flash-medium",
          "verified_publish": true,
          "levels_completed": 1,
          "actions": 552,
          "resets": 4,
          "state": "GAME_OVER",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            115,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            42,
            510,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            59,
            102,
            67,
            98,
            324,
            578
          ]
        },
        {
          "session_id": "fe965bc7-54df-46c0-b57c-58c431cc366b",
          "game_id": "dc22",
          "game_version": "fdcac232",
          "published_at": "2026-08-11T19:40:19.298571Z",
          "score": 0.2273825854072768,
          "model": "grok-4.6",
          "runner": "benchmark_agent",
          "config": "xai-grok-4-6-xhigh",
          "verified_publish": true,
          "levels_completed": 1,
          "actions": 659,
          "resets": 4,
          "state": "GAME_OVER",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            4.775034293552813,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            270,
            389,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            59,
            102,
            67,
            98,
            324,
            578
          ]
        },
        {
          "session_id": "70a3565a-fecc-4a92-8581-5b89df295309",
          "game_id": "dc22",
          "game_version": "fdcac232",
          "published_at": "2026-07-22T12:58:38.398841Z",
          "score": 44.74619047619048,
          "model": "claude-opus-5",
          "runner": "benchmark_agent",
          "config": "anthropic-claude-opus-5-high",
          "verified_publish": true,
          "levels_completed": 4,
          "actions": 563,
          "resets": 2,
          "state": "NOT_FINISHED",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            115,
            115,
            44.89000000000001,
            115,
            0,
            0
          ],
          "level_actions": [
            47,
            48,
            100,
            82,
            286,
            0
          ],
          "level_baseline_actions": [
            59,
            102,
            67,
            98,
            324,
            578
          ]
        },
        {
          "session_id": "1419362a-4433-4fcf-b7a3-2246e9645f85",
          "game_id": "dc22",
          "game_version": "fdcac232",
          "published_at": "2026-06-24T11:05:00.617365Z",
          "score": 0,
          "model": "gpt-5-6-luna",
          "runner": "benchmark_agent",
          "config": "openai-gpt-5-6-luna-low",
          "verified_publish": true,
          "levels_completed": 0,
          "actions": 295,
          "resets": 20,
          "state": "NOT_FINISHED",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            295,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            59,
            102,
            67,
            98,
            324,
            578
          ]
        },
        {
          "session_id": "456323b7-ce30-4b51-a4b1-72c4b24d7bac",
          "game_id": "dc22",
          "game_version": "fdcac232",
          "published_at": "2026-06-24T10:55:27.655633Z",
          "score": 0,
          "model": "gpt-5-6-luna",
          "runner": "benchmark_agent",
          "config": "openai-gpt-5-6-luna-medium",
          "verified_publish": true,
          "levels_completed": 0,
          "actions": 295,
          "resets": 3,
          "state": "GAME_OVER",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            295,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            59,
            102,
            67,
            98,
            324,
            578
          ]
        },
        {
          "session_id": "4ed3a097-d2f4-43d3-9c7d-3ed6e5c58401",
          "game_id": "dc22",
          "game_version": "fdcac232",
          "published_at": "2026-06-24T05:10:48.461802Z",
          "score": 0,
          "model": "gpt-5-6-luna",
          "runner": "benchmark_agent",
          "config": "openai-gpt-5-6-luna-high",
          "verified_publish": true,
          "levels_completed": 0,
          "actions": 295,
          "resets": 7,
          "state": "NOT_FINISHED",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            295,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            59,
            102,
            67,
            98,
            324,
            578
          ]
        },
        {
          "session_id": "228800a7-ed9c-45c4-aac9-ce25e26f35a6",
          "game_id": "dc22",
          "game_version": "fdcac232",
          "published_at": "2026-06-24T04:41:30.08333Z",
          "score": 0,
          "model": "gpt-5-6-luna",
          "runner": "benchmark_agent",
          "config": "openai-gpt-5-6-luna-xhigh",
          "verified_publish": true,
          "levels_completed": 0,
          "actions": 295,
          "resets": 47,
          "state": "NOT_FINISHED",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            295,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            59,
            102,
            67,
            98,
            324,
            578
          ]
        },
        {
          "session_id": "7f9c9815-3879-4722-8d63-6d74a4d4df29",
          "game_id": "dc22",
          "game_version": "fdcac232",
          "published_at": "2026-06-23T04:25:34.90121Z",
          "score": 0,
          "model": "gpt-5-6-luna",
          "runner": "benchmark_agent",
          "config": "openai-gpt-5-6-luna-max",
          "verified_publish": true,
          "levels_completed": 0,
          "actions": 295,
          "resets": 50,
          "state": "NOT_FINISHED",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            295,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            59,
            102,
            67,
            98,
            324,
            578
          ]
        },
        {
          "session_id": "7653d74d-62d0-4636-9693-dc1b0fbf9309",
          "game_id": "dc22",
          "game_version": "fdcac232",
          "published_at": "2026-06-20T12:13:40.282174Z",
          "score": 0,
          "model": "gpt-5-6-terra",
          "runner": "benchmark_agent",
          "config": "openai-gpt-5-6-terra-low",
          "verified_publish": true,
          "levels_completed": 0,
          "actions": 295,
          "resets": 2,
          "state": "GAME_OVER",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            295,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            59,
            102,
            67,
            98,
            324,
            578
          ]
        },
        {
          "session_id": "a9485072-7e5a-4f2e-a433-44ba8fb7e5c1",
          "game_id": "dc22",
          "game_version": "fdcac232",
          "published_at": "2026-06-20T10:08:48.875491Z",
          "score": 0,
          "model": "gpt-5-6-terra",
          "runner": "benchmark_agent",
          "config": "openai-gpt-5-6-terra-high",
          "verified_publish": true,
          "levels_completed": 0,
          "actions": 295,
          "resets": 3,
          "state": "GAME_OVER",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            295,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            59,
            102,
            67,
            98,
            324,
            578
          ]
        },
        {
          "session_id": "61a9d4d5-40af-4667-bb86-5c9ee5d0eeec",
          "game_id": "dc22",
          "game_version": "fdcac232",
          "published_at": "2026-06-20T09:51:42.20915Z",
          "score": 0,
          "model": "gpt-5-6-terra",
          "runner": "benchmark_agent",
          "config": "openai-gpt-5-6-terra-medium",
          "verified_publish": true,
          "levels_completed": 0,
          "actions": 295,
          "resets": 2,
          "state": "GAME_OVER",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            295,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            59,
            102,
            67,
            98,
            324,
            578
          ]
        },
        {
          "session_id": "d5023cf2-9dd1-46d8-8ace-05a96fb85fbe",
          "game_id": "dc22",
          "game_version": "fdcac232",
          "published_at": "2026-06-20T05:32:30.538404Z",
          "score": 0,
          "model": "gpt-5-6-terra",
          "runner": "benchmark_agent",
          "config": "openai-gpt-5-6-terra-xhigh",
          "verified_publish": true,
          "levels_completed": 0,
          "actions": 295,
          "resets": 8,
          "state": "NOT_FINISHED",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            295,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            59,
            102,
            67,
            98,
            324,
            578
          ]
        },
        {
          "session_id": "faa3bed5-c3cc-403d-87e2-740474c8fb92",
          "game_id": "dc22",
          "game_version": "fdcac232",
          "published_at": "2026-06-19T01:27:32.346852Z",
          "score": 0,
          "model": "gpt-5-6-terra",
          "runner": "benchmark_agent",
          "config": "openai-gpt-5-6-terra-max",
          "verified_publish": true,
          "levels_completed": 0,
          "actions": 295,
          "resets": 15,
          "state": "NOT_FINISHED",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            295,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            59,
            102,
            67,
            98,
            324,
            578
          ]
        },
        {
          "session_id": "be1f5f5c-ed6e-46d0-81ec-feff1366a5cd",
          "game_id": "dc22",
          "game_version": "fdcac232",
          "published_at": "2026-06-13T07:30:18.316601Z",
          "score": 14.285714285714285,
          "model": "gpt-5-6-sol",
          "runner": "benchmark_agent",
          "config": "openai-gpt-5-6-sol-max",
          "verified_publish": true,
          "levels_completed": 2,
          "actions": 427,
          "resets": 2,
          "state": "GAME_OVER",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            115,
            115,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            26,
            66,
            335,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            59,
            102,
            67,
            98,
            324,
            578
          ]
        },
        {
          "session_id": "75a31019-6f93-4efd-aa63-40e6ad14ca2f",
          "game_id": "dc22",
          "game_version": "fdcac232",
          "published_at": "2026-06-12T00:31:48.392639Z",
          "score": 1.5035093402440343,
          "model": "gpt-5-6-sol",
          "runner": "benchmark_agent",
          "config": "openai-gpt-5-6-sol-xhigh",
          "verified_publish": true,
          "levels_completed": 1,
          "actions": 343,
          "resets": 3,
          "state": "GAME_OVER",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            31.573696145124718,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            105,
            238,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            59,
            102,
            67,
            98,
            324,
            578
          ]
        },
        {
          "session_id": "14c47878-93ab-43ce-bf0b-eb5d15985909",
          "game_id": "dc22",
          "game_version": "fdcac232",
          "published_at": "2026-06-10T16:23:47.29731Z",
          "score": 2.9468783068783067,
          "model": "gpt-5-6-sol",
          "runner": "benchmark_agent",
          "config": "openai-gpt-5-6-sol-high",
          "verified_publish": true,
          "levels_completed": 1,
          "actions": 585,
          "resets": 6,
          "state": "GAME_OVER",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            61.88444444444444,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            75,
            510,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            59,
            102,
            67,
            98,
            324,
            578
          ]
        },
        {
          "session_id": "444c745a-0f24-48e9-b6aa-6b0198362ae5",
          "game_id": "dc22",
          "game_version": "fdcac232",
          "published_at": "2026-06-10T10:50:56.797456Z",
          "score": 0,
          "model": "gpt-5-6-sol",
          "runner": "benchmark_agent",
          "config": "openai-gpt-5-6-sol-medium",
          "verified_publish": true,
          "levels_completed": 0,
          "actions": 295,
          "resets": 7,
          "state": "NOT_FINISHED",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            295,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            59,
            102,
            67,
            98,
            324,
            578
          ]
        },
        {
          "session_id": "5e6f4e2e-8ab9-4bf8-ba13-37b5421afd10",
          "game_id": "dc22",
          "game_version": "fdcac232",
          "published_at": "2026-06-09T23:07:09.421994Z",
          "score": 0,
          "model": "gpt-5-6-sol",
          "runner": "benchmark_agent",
          "config": "openai-gpt-5-6-sol-low",
          "verified_publish": true,
          "levels_completed": 0,
          "actions": 295,
          "resets": 2,
          "state": "NOT_FINISHED",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            295,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            59,
            102,
            67,
            98,
            324,
            578
          ]
        },
        {
          "session_id": "341070a0-af14-4a2c-84a7-4c417d947aee",
          "game_id": "dc22",
          "game_version": "fdcac232",
          "published_at": "2026-05-30T12:31:36.731647Z",
          "score": 18.11461412561395,
          "model": "claude-opus-4-8",
          "runner": "benchmark_agent",
          "config": "anthropic-opus-4-8-high",
          "verified_publish": true,
          "levels_completed": 3,
          "actions": 866,
          "resets": 25,
          "state": "NOT_FINISHED",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            115,
            16.126232252464508,
            77.71814404432132,
            0,
            0,
            0
          ],
          "level_actions": [
            46,
            254,
            76,
            490,
            0,
            0
          ],
          "level_baseline_actions": [
            59,
            102,
            67,
            98,
            324,
            578
          ]
        },
        {
          "session_id": "1cf5bc80-2cb9-4728-80c9-92573e0c03b3",
          "game_id": "dc22",
          "game_version": "fdcac232",
          "published_at": "2026-05-01T05:09:13.22755Z",
          "score": 0,
          "model": "claude-opus-4-7",
          "runner": "benchmark_agent",
          "config": "anthropic-opus-4-7-high",
          "verified_publish": true,
          "levels_completed": 0,
          "actions": 295,
          "resets": 16,
          "state": "NOT_FINISHED",
          "completed": false,
          "number_of_levels": 6,
          "level_scores": [
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            295,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            59,
            102,
            67,
            98,
            324,
            578
          ]
        },
        {
          "session_id": "7369f85a-4b3e-4292-8c65-bc59c34e0787",
          "game_id": "dc22",
          "game_version": "fdcac232",
          "published_at": "2026-04-30T16:48:58.904762Z",
          "score": 0,
          "model": "gpt-5.5-2026-04-23",
          "runner": "benchmark_agent",
          "config": "openai-gpt-5-5-2026-04-23-high",
          "verified_publish": true,
          "levels_completed": 0,
          "actions": 295,
          "resets": 6,
          "state": "GAME_OVER",
          "completed": false,
          "number_of_levels": 6,
          "level_scores": [
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            295,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            59,
            102,
            67,
            98,
            324,
            578
          ]
        },
        {
          "session_id": "b5be0b28-d326-4ff4-9efa-aa86e1cc3c79",
          "game_id": "dc22",
          "game_version": "fdcac232",
          "published_at": "2026-04-14T17:43:37.196834Z",
          "score": 0,
          "model": "grok-4.20-beta-0309-reasoning",
          "runner": "benchmark_agent",
          "config": "xai-grok-4-20-beta-0309-reasoning",
          "verified_publish": true,
          "levels_completed": 0,
          "actions": 295,
          "resets": 15,
          "state": "NOT_FINISHED",
          "completed": false,
          "number_of_levels": 6,
          "level_scores": [
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            295,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            59,
            102,
            67,
            98,
            324,
            578
          ]
        },
        {
          "session_id": "c76a9be7-48fe-45d2-92b0-a7c970a5d436",
          "game_id": "dc22",
          "game_version": "fdcac232",
          "published_at": "2026-04-14T17:43:36.851531Z",
          "score": 0,
          "model": "gpt-5.4-2026-03-05",
          "runner": "benchmark_agent",
          "config": "openai-gpt-5-4-2026-03-05-high",
          "verified_publish": true,
          "levels_completed": 0,
          "actions": 295,
          "resets": 2,
          "state": "GAME_OVER",
          "completed": false,
          "number_of_levels": 6,
          "level_scores": [
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            295,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            59,
            102,
            67,
            98,
            324,
            578
          ]
        },
        {
          "session_id": "5c57fea0-8b98-4ae8-9fd7-1148b67e3459",
          "game_id": "dc22",
          "game_version": "fdcac232",
          "published_at": "2026-04-14T17:43:36.5395Z",
          "score": 0,
          "model": "gemini-3.1-pro-preview",
          "runner": "benchmark_agent",
          "config": "google-gemini-3-1-pro-preview",
          "verified_publish": true,
          "levels_completed": 0,
          "actions": 100,
          "resets": 2,
          "state": "GAME_OVER",
          "completed": false,
          "number_of_levels": 6,
          "level_scores": [
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            100,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            59,
            102,
            67,
            98,
            324,
            578
          ]
        },
        {
          "session_id": "42384808-f7f1-47be-a26f-5f341a49c4d1",
          "game_id": "dc22",
          "game_version": "fdcac232",
          "published_at": "2026-04-14T17:43:36.097267Z",
          "score": 0,
          "model": "claude-opus-4-6",
          "runner": "benchmark_agent",
          "config": "anthropic-opus-4-6-max-effort",
          "verified_publish": true,
          "levels_completed": 0,
          "actions": 295,
          "resets": 19,
          "state": "GAME_OVER",
          "completed": false,
          "number_of_levels": 6,
          "level_scores": [
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            295,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            59,
            102,
            67,
            98,
            324,
            578
          ]
        }
      ]
    },
    {
      "game": "wa30",
      "source": "https://arcprize.org/tasks/wa30",
      "runs": [
        {
          "session_id": "86b92065-fed4-4dc7-94a9-dc9c25d34740",
          "game_id": "wa30",
          "game_version": "ee6fef47",
          "published_at": "2026-09-25T15:24:41.448344Z",
          "score": 0.1228259969104669,
          "model": "gpt-6-sol",
          "runner": "benchmark_agent",
          "config": "openai-gpt-6-sol-max",
          "verified_publish": true,
          "levels_completed": 1,
          "actions": 443,
          "resets": 4,
          "state": "GAME_OVER",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            5.527169860971011,
            0,
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            302,
            141,
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            71,
            119,
            183,
            98,
            368,
            68,
            79,
            442,
            415
          ]
        },
        {
          "session_id": "2aeeff31-68ed-4616-a957-7ad3d4534cea",
          "game_id": "wa30",
          "game_version": "ee6fef47",
          "published_at": "2026-09-25T07:35:38.892841Z",
          "score": 5.72577351013565,
          "model": "gpt-6-sol",
          "runner": "benchmark_agent",
          "config": "openai-gpt-6-sol-low-provider-adapter",
          "verified_publish": true,
          "levels_completed": 2,
          "actions": 1109,
          "resets": 10,
          "state": "GAME_OVER",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            27.65980795610425,
            115,
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            135,
            59,
            915,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            71,
            119,
            183,
            98,
            368,
            68,
            79,
            442,
            415
          ]
        },
        {
          "session_id": "f2a5e15b-ae5a-46e9-839e-6d2ba9c469a7",
          "game_id": "wa30",
          "game_version": "ee6fef47",
          "published_at": "2026-09-25T07:17:26.041807Z",
          "score": 0.15832186983750104,
          "model": "gpt-6-sol",
          "runner": "benchmark_agent",
          "config": "openai-gpt-6-sol-none-provider-adapter",
          "verified_publish": true,
          "levels_completed": 1,
          "actions": 861,
          "resets": 10,
          "state": "GAME_OVER",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            7.124484142687547,
            0,
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            266,
            595,
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            71,
            119,
            183,
            98,
            368,
            68,
            79,
            442,
            415
          ]
        },
        {
          "session_id": "2ec382dc-432d-476d-9be4-45cc09ebdd20",
          "game_id": "wa30",
          "game_version": "ee6fef47",
          "published_at": "2026-09-25T07:04:41.062339Z",
          "score": 8.282719469615861,
          "model": "gpt-6-sol",
          "runner": "benchmark_agent",
          "config": "openai-gpt-6-sol-xhigh-provider-adapter",
          "verified_publish": true,
          "levels_completed": 3,
          "actions": 1185,
          "resets": 9,
          "state": "GAME_OVER",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            115,
            115,
            9.240792044237923,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            45,
            48,
            602,
            490,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            71,
            119,
            183,
            98,
            368,
            68,
            79,
            442,
            415
          ]
        },
        {
          "session_id": "20fb9bc9-ed7f-45a4-96ac-b08a6e4d346b",
          "game_id": "wa30",
          "game_version": "ee6fef47",
          "published_at": "2026-09-25T06:46:10.629948Z",
          "score": 0.1459972399252202,
          "model": "gpt-6-sol",
          "runner": "benchmark_agent",
          "config": "openai-gpt-6-sol-medium-provider-adapter",
          "verified_publish": true,
          "levels_completed": 1,
          "actions": 872,
          "resets": 9,
          "state": "GAME_OVER",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            6.56987579663491,
            0,
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            277,
            595,
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            71,
            119,
            183,
            98,
            368,
            68,
            79,
            442,
            415
          ]
        },
        {
          "session_id": "a025225d-d06f-482b-9dfc-8df3e56b5efe",
          "game_id": "wa30",
          "game_version": "ee6fef47",
          "published_at": "2026-09-25T06:37:28.044731Z",
          "score": 0,
          "model": "gpt-6-sol",
          "runner": "benchmark_agent",
          "config": "openai-gpt-6-sol-none",
          "verified_publish": true,
          "levels_completed": 0,
          "actions": 355,
          "resets": 2,
          "state": "NOT_FINISHED",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            0,
            0,
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            355,
            0,
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            71,
            119,
            183,
            98,
            368,
            68,
            79,
            442,
            415
          ]
        },
        {
          "session_id": "9f5ac032-f644-4813-89aa-53f5ff7e8bd0",
          "game_id": "wa30",
          "game_version": "ee6fef47",
          "published_at": "2026-09-25T06:36:32.8237Z",
          "score": 6.666666666666667,
          "model": "gpt-6-sol",
          "runner": "benchmark_agent",
          "config": "openai-gpt-6-sol-high-provider-adapter",
          "verified_publish": true,
          "levels_completed": 2,
          "actions": 1038,
          "resets": 9,
          "state": "GAME_OVER",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            97.24151234567903,
            115,
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            72,
            51,
            915,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            71,
            119,
            183,
            98,
            368,
            68,
            79,
            442,
            415
          ]
        },
        {
          "session_id": "c7f5c616-1a3a-4d92-abc7-398d00276dcc",
          "game_id": "wa30",
          "game_version": "ee6fef47",
          "published_at": "2026-09-25T06:19:24.926862Z",
          "score": 0,
          "model": "gpt-6-sol",
          "runner": "benchmark_agent",
          "config": "openai-gpt-6-sol-low",
          "verified_publish": true,
          "levels_completed": 0,
          "actions": 355,
          "resets": 4,
          "state": "NOT_FINISHED",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            0,
            0,
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            355,
            0,
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            71,
            119,
            183,
            98,
            368,
            68,
            79,
            442,
            415
          ]
        },
        {
          "session_id": "7c14a998-c92d-4f70-a8b5-698fe594424c",
          "game_id": "wa30",
          "game_version": "ee6fef47",
          "published_at": "2026-09-25T06:16:05.083908Z",
          "score": 0,
          "model": "gpt-6-sol",
          "runner": "benchmark_agent",
          "config": "openai-gpt-6-sol-xhigh",
          "verified_publish": true,
          "levels_completed": 0,
          "actions": 355,
          "resets": 4,
          "state": "NOT_FINISHED",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            0,
            0,
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            355,
            0,
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            71,
            119,
            183,
            98,
            368,
            68,
            79,
            442,
            415
          ]
        },
        {
          "session_id": "d815c9bb-edaf-42bd-b0a6-7611fef7f497",
          "game_id": "wa30",
          "game_version": "ee6fef47",
          "published_at": "2026-09-25T05:58:37.480923Z",
          "score": 0,
          "model": "gpt-6-sol",
          "runner": "benchmark_agent",
          "config": "openai-gpt-6-sol-medium",
          "verified_publish": true,
          "levels_completed": 0,
          "actions": 355,
          "resets": 3,
          "state": "NOT_FINISHED",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            0,
            0,
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            355,
            0,
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            71,
            119,
            183,
            98,
            368,
            68,
            79,
            442,
            415
          ]
        },
        {
          "session_id": "171c044f-dffa-4812-b9d7-340bb8e0d1f5",
          "game_id": "wa30",
          "game_version": "ee6fef47",
          "published_at": "2026-09-25T05:24:20.387706Z",
          "score": 0,
          "model": "gpt-6-sol",
          "runner": "benchmark_agent",
          "config": "openai-gpt-6-sol-high",
          "verified_publish": true,
          "levels_completed": 0,
          "actions": 355,
          "resets": 2,
          "state": "NOT_FINISHED",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            0,
            0,
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            355,
            0,
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            71,
            119,
            183,
            98,
            368,
            68,
            79,
            442,
            415
          ]
        },
        {
          "session_id": "e68818c8-ce9c-4813-b32b-51212dc02586",
          "game_id": "wa30",
          "game_version": "ee6fef47",
          "published_at": "2026-09-21T18:53:25.070154Z",
          "score": 13.333333333333334,
          "model": "gemini-3.8-flash",
          "runner": "benchmark_agent",
          "config": "google-gemini-3-8-flash-low-provider-adapter",
          "verified_publish": true,
          "levels_completed": 3,
          "actions": 675,
          "resets": 6,
          "state": "GAME_OVER",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            115,
            115,
            115,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            39,
            66,
            80,
            490,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            71,
            119,
            183,
            98,
            368,
            68,
            79,
            442,
            415
          ]
        },
        {
          "session_id": "fbc54dc3-3909-418e-81b0-7a20e3f4945a",
          "game_id": "wa30",
          "game_version": "ee6fef47",
          "published_at": "2026-09-20T17:44:22.634711Z",
          "score": 14.890941332969605,
          "model": "gemini-3.8-flash",
          "runner": "benchmark_agent",
          "config": "google-gemini-3-8-flash-medium-provider-adapter",
          "verified_publish": true,
          "levels_completed": 4,
          "actions": 2392,
          "resets": 17,
          "state": "GAME_OVER",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            115,
            82.51850125284074,
            115,
            11.263839369487709,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            41,
            131,
            88,
            292,
            1840,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            71,
            119,
            183,
            98,
            368,
            68,
            79,
            442,
            415
          ]
        },
        {
          "session_id": "c9c874f7-e724-49c0-905c-bb61e760ca59",
          "game_id": "wa30",
          "game_version": "ee6fef47",
          "published_at": "2026-09-18T21:54:49.734281Z",
          "score": 43.282609212928136,
          "model": "gemini-3.8-flash",
          "runner": "benchmark_agent",
          "config": "google-gemini-3-8-flash-high-provider-adapter",
          "verified_publish": true,
          "levels_completed": 7,
          "actions": 2556,
          "resets": 14,
          "state": "GAME_OVER",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            115,
            115,
            115,
            115,
            115,
            23.591836734693878,
            11.59519916765755,
            0,
            0
          ],
          "level_actions": [
            41,
            70,
            76,
            57,
            107,
            140,
            232,
            1833,
            0
          ],
          "level_baseline_actions": [
            71,
            119,
            183,
            98,
            368,
            68,
            79,
            442,
            415
          ]
        },
        {
          "session_id": "c9e1f2c7-9118-4811-995c-7532c5fbcebc",
          "game_id": "wa30",
          "game_version": "ee6fef47",
          "published_at": "2026-09-16T19:58:10.175462Z",
          "score": 0,
          "model": "gpt-6-luna",
          "runner": "benchmark_agent",
          "config": "openai-gpt-6-luna-medium",
          "verified_publish": true,
          "levels_completed": 0,
          "actions": 355,
          "resets": 5,
          "state": "NOT_FINISHED",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            0,
            0,
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            355,
            0,
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            71,
            119,
            183,
            98,
            368,
            68,
            79,
            442,
            415
          ]
        },
        {
          "session_id": "41ad192e-c73c-47ee-808c-2740a9a3d42f",
          "game_id": "wa30",
          "game_version": "ee6fef47",
          "published_at": "2026-09-16T19:44:37.366014Z",
          "score": 0,
          "model": "gpt-6-luna",
          "runner": "benchmark_agent",
          "config": "openai-gpt-6-luna-medium-provider-adapter",
          "verified_publish": true,
          "levels_completed": 0,
          "actions": 355,
          "resets": 2,
          "state": "NOT_FINISHED",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            0,
            0,
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            355,
            0,
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            71,
            119,
            183,
            98,
            368,
            68,
            79,
            442,
            415
          ]
        },
        {
          "session_id": "e9bd37af-c451-438c-ad6c-db7aab194c1f",
          "game_id": "wa30",
          "game_version": "ee6fef47",
          "published_at": "2026-09-16T19:40:13.457518Z",
          "score": 0,
          "model": "gpt-6-luna",
          "runner": "benchmark_agent",
          "config": "openai-gpt-6-luna-low",
          "verified_publish": true,
          "levels_completed": 0,
          "actions": 355,
          "resets": 9,
          "state": "NOT_FINISHED",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            0,
            0,
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            355,
            0,
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            71,
            119,
            183,
            98,
            368,
            68,
            79,
            442,
            415
          ]
        },
        {
          "session_id": "492dbb9f-b6e8-4ef0-988a-0bb9b604cc06",
          "game_id": "wa30",
          "game_version": "ee6fef47",
          "published_at": "2026-09-16T19:36:00.238304Z",
          "score": 0,
          "model": "gpt-6-luna",
          "runner": "benchmark_agent",
          "config": "openai-gpt-6-luna-low-provider-adapter",
          "verified_publish": true,
          "levels_completed": 0,
          "actions": 355,
          "resets": 2,
          "state": "GAME_OVER",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            0,
            0,
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            355,
            0,
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            71,
            119,
            183,
            98,
            368,
            68,
            79,
            442,
            415
          ]
        },
        {
          "session_id": "c42724e6-51e2-4b43-a1a8-3ecbf7056d30",
          "game_id": "wa30",
          "game_version": "ee6fef47",
          "published_at": "2026-09-16T19:33:29.637885Z",
          "score": 0,
          "model": "gpt-6-luna",
          "runner": "benchmark_agent",
          "config": "openai-gpt-6-luna-none-provider-adapter",
          "verified_publish": true,
          "levels_completed": 0,
          "actions": 355,
          "resets": 8,
          "state": "NOT_FINISHED",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            0,
            0,
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            355,
            0,
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            71,
            119,
            183,
            98,
            368,
            68,
            79,
            442,
            415
          ]
        },
        {
          "session_id": "95b7ae93-8d56-41c8-8851-7254e41184b2",
          "game_id": "wa30",
          "game_version": "ee6fef47",
          "published_at": "2026-09-16T19:30:37.783476Z",
          "score": 0,
          "model": "gpt-6-luna",
          "runner": "benchmark_agent",
          "config": "openai-gpt-6-luna-none",
          "verified_publish": true,
          "levels_completed": 0,
          "actions": 355,
          "resets": 23,
          "state": "NOT_FINISHED",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            0,
            0,
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            355,
            0,
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            71,
            119,
            183,
            98,
            368,
            68,
            79,
            442,
            415
          ]
        },
        {
          "session_id": "9a7ab982-fa9c-4ce1-b3c3-e9a040564cc8",
          "game_id": "wa30",
          "game_version": "ee6fef47",
          "published_at": "2026-09-16T17:33:36.975989Z",
          "score": 38.5627601128355,
          "model": "gpt-6-1-sol",
          "runner": "benchmark_agent",
          "config": "openai-gpt-6-1-sol-high",
          "verified_publish": true,
          "levels_completed": 6,
          "actions": 1061,
          "resets": 6,
          "state": "GAME_OVER",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            115,
            115,
            115,
            46.31558641975309,
            34.012371879716994,
            115,
            0,
            0,
            0
          ],
          "level_actions": [
            27,
            53,
            78,
            144,
            631,
            48,
            80,
            0,
            0
          ],
          "level_baseline_actions": [
            71,
            119,
            183,
            98,
            368,
            68,
            79,
            442,
            415
          ]
        },
        {
          "session_id": "e787aa4b-5ccb-4e94-b770-8db8c1b50bdf",
          "game_id": "wa30",
          "game_version": "ee6fef47",
          "published_at": "2026-09-16T16:43:03.931157Z",
          "score": 0,
          "model": "gpt-6-luna",
          "runner": "benchmark_agent",
          "config": "openai-gpt-6-luna-high",
          "verified_publish": true,
          "levels_completed": 0,
          "actions": 355,
          "resets": 5,
          "state": "GAME_OVER",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            0,
            0,
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            355,
            0,
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            71,
            119,
            183,
            98,
            368,
            68,
            79,
            442,
            415
          ]
        },
        {
          "session_id": "53fd1a19-24a7-4b77-a695-00d9d0f35c2c",
          "game_id": "wa30",
          "game_version": "ee6fef47",
          "published_at": "2026-09-16T15:32:18.353436Z",
          "score": 76.42121752924847,
          "model": "gpt-6-1-sol",
          "runner": "benchmark_agent",
          "config": "openai-gpt-6-1-sol-xhigh",
          "verified_publish": true,
          "levels_completed": 8,
          "actions": 2065,
          "resets": 26,
          "state": "GAME_OVER",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            115,
            115,
            115,
            115,
            115,
            115,
            14.850684116597265,
            115,
            0
          ],
          "level_actions": [
            29,
            53,
            76,
            80,
            118,
            50,
            205,
            137,
            1317
          ],
          "level_baseline_actions": [
            71,
            119,
            183,
            98,
            368,
            68,
            79,
            442,
            415
          ]
        },
        {
          "session_id": "65e10ad5-1ab4-43a1-aaa4-a3f69d48f50b",
          "game_id": "wa30",
          "game_version": "ee6fef47",
          "published_at": "2026-09-16T12:22:21.39233Z",
          "score": 62.22222222222222,
          "model": "gpt-6-1-sol",
          "runner": "benchmark_agent",
          "config": "openai-gpt-6-1-sol-max",
          "verified_publish": true,
          "levels_completed": 7,
          "actions": 882,
          "resets": 3,
          "state": "GAME_OVER",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            115,
            115,
            115,
            115,
            115,
            115,
            115,
            0,
            0
          ],
          "level_actions": [
            27,
            49,
            80,
            48,
            122,
            48,
            48,
            460,
            0
          ],
          "level_baseline_actions": [
            71,
            119,
            183,
            98,
            368,
            68,
            79,
            442,
            415
          ]
        },
        {
          "session_id": "b37096da-265a-4a10-9730-8cd01547aaa7",
          "game_id": "wa30",
          "game_version": "ee6fef47",
          "published_at": "2026-09-16T08:55:50.92018Z",
          "score": 0,
          "model": "gpt-6-luna",
          "runner": "benchmark_agent",
          "config": "openai-gpt-6-luna-high-provider-adapter",
          "verified_publish": true,
          "levels_completed": 0,
          "actions": 355,
          "resets": 1,
          "state": "GAME_OVER",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            0,
            0,
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            355,
            0,
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            71,
            119,
            183,
            98,
            368,
            68,
            79,
            442,
            415
          ]
        },
        {
          "session_id": "b9b7dd49-ff4d-4836-a5ac-aa8ea6f39a7c",
          "game_id": "wa30",
          "game_version": "ee6fef47",
          "published_at": "2026-09-16T08:47:21.558872Z",
          "score": 0,
          "model": "gpt-6-luna",
          "runner": "benchmark_agent",
          "config": "openai-gpt-6-luna-xhigh",
          "verified_publish": true,
          "levels_completed": 0,
          "actions": 355,
          "resets": 8,
          "state": "GAME_OVER",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            0,
            0,
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            355,
            0,
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            71,
            119,
            183,
            98,
            368,
            68,
            79,
            442,
            415
          ]
        },
        {
          "session_id": "495f740e-01b8-4f56-8a9d-e6dfd87fd552",
          "game_id": "wa30",
          "game_version": "ee6fef47",
          "published_at": "2026-09-16T08:39:31.647466Z",
          "score": 0,
          "model": "gpt-6-luna",
          "runner": "benchmark_agent",
          "config": "openai-gpt-6-luna-max",
          "verified_publish": true,
          "levels_completed": 0,
          "actions": 355,
          "resets": 12,
          "state": "NOT_FINISHED",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            0,
            0,
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            355,
            0,
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            71,
            119,
            183,
            98,
            368,
            68,
            79,
            442,
            415
          ]
        },
        {
          "session_id": "dcbd3566-d4e9-47e8-91df-ba067608bea4",
          "game_id": "wa30",
          "game_version": "ee6fef47",
          "published_at": "2026-09-16T08:11:46.130241Z",
          "score": 0,
          "model": "gpt-6-luna",
          "runner": "benchmark_agent",
          "config": "openai-gpt-6-luna-xhigh-provider-adapter",
          "verified_publish": true,
          "levels_completed": 0,
          "actions": 355,
          "resets": 1,
          "state": "NOT_FINISHED",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            0,
            0,
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            355,
            0,
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            71,
            119,
            183,
            98,
            368,
            68,
            79,
            442,
            415
          ]
        },
        {
          "session_id": "e4e9c16e-1f74-4337-8356-6ec2b8fdd130",
          "game_id": "wa30",
          "game_version": "ee6fef47",
          "published_at": "2026-09-16T07:48:55.763513Z",
          "score": 6.666666666666667,
          "model": "gpt-6-1-sol",
          "runner": "benchmark_agent",
          "config": "openai-gpt-6-1-sol-low",
          "verified_publish": true,
          "levels_completed": 2,
          "actions": 1008,
          "resets": 11,
          "state": "GAME_OVER",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            115,
            115,
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            38,
            55,
            915,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            71,
            119,
            183,
            98,
            368,
            68,
            79,
            442,
            415
          ]
        },
        {
          "session_id": "da303391-ed67-46b4-a074-31605183a13e",
          "game_id": "wa30",
          "game_version": "ee6fef47",
          "published_at": "2026-09-16T07:44:13.317147Z",
          "score": 0,
          "model": "gpt-6-luna",
          "runner": "benchmark_agent",
          "config": "openai-gpt-6-luna-max-provider-adapter",
          "verified_publish": true,
          "levels_completed": 0,
          "actions": 355,
          "resets": 1,
          "state": "GAME_OVER",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            0,
            0,
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            355,
            0,
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            71,
            119,
            183,
            98,
            368,
            68,
            79,
            442,
            415
          ]
        },
        {
          "session_id": "41e90b4f-3bab-4835-bf67-c0629ba27b52",
          "game_id": "wa30",
          "game_version": "ee6fef47",
          "published_at": "2026-09-16T06:02:40.8345Z",
          "score": 13.333333333333334,
          "model": "gpt-6-1-sol",
          "runner": "benchmark_agent",
          "config": "openai-gpt-6-1-sol-medium",
          "verified_publish": true,
          "levels_completed": 3,
          "actions": 666,
          "resets": 5,
          "state": "GAME_OVER",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            115,
            115,
            115,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            28,
            53,
            95,
            490,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            71,
            119,
            183,
            98,
            368,
            68,
            79,
            442,
            415
          ]
        },
        {
          "session_id": "ee2b0c32-fd88-4c1c-9f20-1fd3daf8afc3",
          "game_id": "wa30",
          "game_version": "ee6fef47",
          "published_at": "2026-09-16T05:24:35.15433Z",
          "score": 100,
          "model": "gpt-6-1-sol",
          "runner": "benchmark_agent",
          "config": "openai-gpt-6-1-sol-high-provider-adapter",
          "verified_publish": true,
          "levels_completed": 9,
          "actions": 696,
          "resets": 1,
          "state": "WIN",
          "completed": true,
          "number_of_levels": 0,
          "level_scores": [
            115,
            115,
            115,
            115,
            115,
            115,
            115,
            115,
            115
          ],
          "level_actions": [
            27,
            63,
            72,
            53,
            118,
            49,
            38,
            139,
            137
          ],
          "level_baseline_actions": [
            71,
            119,
            183,
            98,
            368,
            68,
            79,
            442,
            415
          ]
        },
        {
          "session_id": "9fb32e2c-3fa3-41a9-8897-0ce536f2c669",
          "game_id": "wa30",
          "game_version": "ee6fef47",
          "published_at": "2026-09-16T04:58:07.818223Z",
          "score": 100,
          "model": "gpt-6-1-sol",
          "runner": "benchmark_agent",
          "config": "openai-gpt-6-1-sol-xhigh-provider-adapter",
          "verified_publish": true,
          "levels_completed": 9,
          "actions": 750,
          "resets": 2,
          "state": "WIN",
          "completed": true,
          "number_of_levels": 0,
          "level_scores": [
            115,
            115,
            115,
            115,
            115,
            115,
            115,
            115,
            115
          ],
          "level_actions": [
            26,
            53,
            74,
            54,
            120,
            46,
            55,
            129,
            193
          ],
          "level_baseline_actions": [
            71,
            119,
            183,
            98,
            368,
            68,
            79,
            442,
            415
          ]
        },
        {
          "session_id": "bc7bcbcc-6830-42f0-bd5d-1b75f8816247",
          "game_id": "wa30",
          "game_version": "ee6fef47",
          "published_at": "2026-09-16T03:39:53.543343Z",
          "score": 100,
          "model": "gpt-6-1-sol",
          "runner": "benchmark_agent",
          "config": "openai-gpt-6-1-sol-max-provider-adapter",
          "verified_publish": true,
          "levels_completed": 9,
          "actions": 714,
          "resets": 2,
          "state": "WIN",
          "completed": true,
          "number_of_levels": 0,
          "level_scores": [
            115,
            115,
            115,
            115,
            115,
            115,
            115,
            115,
            115
          ],
          "level_actions": [
            27,
            48,
            75,
            53,
            109,
            47,
            38,
            110,
            207
          ],
          "level_baseline_actions": [
            71,
            119,
            183,
            98,
            368,
            68,
            79,
            442,
            415
          ]
        },
        {
          "session_id": "86ed7fef-5885-4d75-a67a-372ff8d56e7d",
          "game_id": "wa30",
          "game_version": "ee6fef47",
          "published_at": "2026-09-12T14:20:46.781605Z",
          "score": 2.2222222222222223,
          "model": "gemini-3.8-flash",
          "runner": "benchmark_agent",
          "config": "google-gemini-3-8-flash-low",
          "verified_publish": true,
          "levels_completed": 1,
          "actions": 655,
          "resets": 8,
          "state": "GAME_OVER",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            115,
            0,
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            60,
            595,
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            71,
            119,
            183,
            98,
            368,
            68,
            79,
            442,
            415
          ]
        },
        {
          "session_id": "49da413c-1191-4c79-babf-1c1f683be447",
          "game_id": "wa30",
          "game_version": "ee6fef47",
          "published_at": "2026-09-11T03:46:54.90626Z",
          "score": 16.090314250271263,
          "model": "gemini-3.8-flash",
          "runner": "benchmark_agent",
          "config": "google-gemini-3-8-flash-high",
          "verified_publish": true,
          "levels_completed": 4,
          "actions": 2172,
          "resets": 17,
          "state": "GAME_OVER",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            115,
            11.364347679541607,
            42.1118153010412,
            115,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            59,
            353,
            282,
            91,
            1387,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            71,
            119,
            183,
            98,
            368,
            68,
            79,
            442,
            415
          ]
        },
        {
          "session_id": "db6a42f8-b176-4434-b3c6-b2e809780e5f",
          "game_id": "wa30",
          "game_version": "ee6fef47",
          "published_at": "2026-09-11T03:23:13.13855Z",
          "score": 2.2222222222222223,
          "model": "gemini-3.8-flash",
          "runner": "benchmark_agent",
          "config": "google-gemini-3-8-flash-medium",
          "verified_publish": true,
          "levels_completed": 1,
          "actions": 664,
          "resets": 8,
          "state": "GAME_OVER",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            105.88111741230833,
            0,
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            69,
            595,
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            71,
            119,
            183,
            98,
            368,
            68,
            79,
            442,
            415
          ]
        },
        {
          "session_id": "e1ed9301-bccd-45b7-ac65-f89f85d0389d",
          "game_id": "wa30",
          "game_version": "ee6fef47",
          "published_at": "2026-08-11T16:31:11.315473Z",
          "score": 6.223044500126255,
          "model": "grok-4.6",
          "runner": "benchmark_agent",
          "config": "xai-grok-4-6-xhigh",
          "verified_publish": true,
          "levels_completed": 2,
          "actions": 914,
          "resets": 8,
          "state": "GAME_OVER",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            115,
            82.51850125284074,
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            31,
            131,
            752,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            71,
            119,
            183,
            98,
            368,
            68,
            79,
            442,
            415
          ]
        },
        {
          "session_id": "086c9f8f-5a86-4582-8517-8edb90b718cd",
          "game_id": "wa30",
          "game_version": "ee6fef47",
          "published_at": "2026-07-21T06:31:40.121403Z",
          "score": 10.597511027095146,
          "model": "claude-opus-5",
          "runner": "benchmark_agent",
          "config": "anthropic-claude-opus-5-high",
          "verified_publish": true,
          "levels_completed": 3,
          "actions": 877,
          "resets": 11,
          "state": "GAME_OVER",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            115,
            115,
            43.96266540642721,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            44,
            67,
            276,
            490,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            71,
            119,
            183,
            98,
            368,
            68,
            79,
            442,
            415
          ]
        },
        {
          "session_id": "52c0f9c6-e8cc-48a7-a147-2d8cc6a94af2",
          "game_id": "wa30",
          "game_version": "ee6fef47",
          "published_at": "2026-06-24T11:00:29.817095Z",
          "score": 0,
          "model": "gpt-5-6-luna",
          "runner": "benchmark_agent",
          "config": "openai-gpt-5-6-luna-low",
          "verified_publish": true,
          "levels_completed": 0,
          "actions": 355,
          "resets": 16,
          "state": "NOT_FINISHED",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            0,
            0,
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            355,
            0,
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            71,
            119,
            183,
            98,
            368,
            68,
            79,
            442,
            415
          ]
        },
        {
          "session_id": "b11665d7-38c3-4179-840f-ced40482311b",
          "game_id": "wa30",
          "game_version": "ee6fef47",
          "published_at": "2026-06-24T10:34:06.963746Z",
          "score": 0,
          "model": "gpt-5-6-luna",
          "runner": "benchmark_agent",
          "config": "openai-gpt-5-6-luna-medium",
          "verified_publish": true,
          "levels_completed": 0,
          "actions": 355,
          "resets": 27,
          "state": "NOT_FINISHED",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            0,
            0,
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            355,
            0,
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            71,
            119,
            183,
            98,
            368,
            68,
            79,
            442,
            415
          ]
        },
        {
          "session_id": "e342b28b-f4c9-4971-b719-c7fd847ca87e",
          "game_id": "wa30",
          "game_version": "ee6fef47",
          "published_at": "2026-06-24T04:50:16.018508Z",
          "score": 0,
          "model": "gpt-5-6-luna",
          "runner": "benchmark_agent",
          "config": "openai-gpt-5-6-luna-high",
          "verified_publish": true,
          "levels_completed": 0,
          "actions": 355,
          "resets": 32,
          "state": "NOT_FINISHED",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            0,
            0,
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            355,
            0,
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            71,
            119,
            183,
            98,
            368,
            68,
            79,
            442,
            415
          ]
        },
        {
          "session_id": "b2409e3c-d05a-4712-84aa-15af781fad5b",
          "game_id": "wa30",
          "game_version": "ee6fef47",
          "published_at": "2026-06-24T04:00:52.008118Z",
          "score": 0,
          "model": "gpt-5-6-luna",
          "runner": "benchmark_agent",
          "config": "openai-gpt-5-6-luna-xhigh",
          "verified_publish": true,
          "levels_completed": 0,
          "actions": 355,
          "resets": 41,
          "state": "NOT_FINISHED",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            0,
            0,
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            355,
            0,
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            71,
            119,
            183,
            98,
            368,
            68,
            79,
            442,
            415
          ]
        },
        {
          "session_id": "43b14b08-78db-4164-8b68-3823f8cab97a",
          "game_id": "wa30",
          "game_version": "ee6fef47",
          "published_at": "2026-06-23T04:52:49.043065Z",
          "score": 0,
          "model": "gpt-5-6-luna",
          "runner": "benchmark_agent",
          "config": "openai-gpt-5-6-luna-max",
          "verified_publish": true,
          "levels_completed": 0,
          "actions": 355,
          "resets": 61,
          "state": "NOT_FINISHED",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            0,
            0,
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            355,
            0,
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            71,
            119,
            183,
            98,
            368,
            68,
            79,
            442,
            415
          ]
        },
        {
          "session_id": "eff719cb-cca1-4ec6-9a7c-dcc154b118f4",
          "game_id": "wa30",
          "game_version": "ee6fef47",
          "published_at": "2026-06-20T11:53:14.894463Z",
          "score": 0,
          "model": "gpt-5-6-terra",
          "runner": "benchmark_agent",
          "config": "openai-gpt-5-6-terra-low",
          "verified_publish": true,
          "levels_completed": 0,
          "actions": 36,
          "resets": 0,
          "state": "NOT_FINISHED",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            0,
            0,
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            36,
            0,
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            71,
            119,
            183,
            98,
            368,
            68,
            79,
            442,
            415
          ]
        },
        {
          "session_id": "1468e31d-67ef-4c28-bd77-972abbc5c768",
          "game_id": "wa30",
          "game_version": "ee6fef47",
          "published_at": "2026-06-20T09:29:53.334517Z",
          "score": 0,
          "model": "gpt-5-6-terra",
          "runner": "benchmark_agent",
          "config": "openai-gpt-5-6-terra-medium",
          "verified_publish": true,
          "levels_completed": 0,
          "actions": 355,
          "resets": 3,
          "state": "NOT_FINISHED",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            0,
            0,
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            355,
            0,
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            71,
            119,
            183,
            98,
            368,
            68,
            79,
            442,
            415
          ]
        },
        {
          "session_id": "55f12462-04b3-44fd-8c48-a17bc573cb3b",
          "game_id": "wa30",
          "game_version": "ee6fef47",
          "published_at": "2026-06-20T08:48:24.799785Z",
          "score": 0,
          "model": "gpt-5-6-terra",
          "runner": "benchmark_agent",
          "config": "openai-gpt-5-6-terra-high",
          "verified_publish": true,
          "levels_completed": 0,
          "actions": 355,
          "resets": 4,
          "state": "NOT_FINISHED",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            0,
            0,
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            355,
            0,
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            71,
            119,
            183,
            98,
            368,
            68,
            79,
            442,
            415
          ]
        },
        {
          "session_id": "28b02c42-285f-4564-918f-79ed32f2d71f",
          "game_id": "wa30",
          "game_version": "ee6fef47",
          "published_at": "2026-06-20T03:15:54.203032Z",
          "score": 0,
          "model": "gpt-5-6-terra",
          "runner": "benchmark_agent",
          "config": "openai-gpt-5-6-terra-xhigh",
          "verified_publish": true,
          "levels_completed": 0,
          "actions": 355,
          "resets": 5,
          "state": "NOT_FINISHED",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            0,
            0,
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            355,
            0,
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            71,
            119,
            183,
            98,
            368,
            68,
            79,
            442,
            415
          ]
        },
        {
          "session_id": "6a0eeddd-b01c-420c-a560-f5f758751b91",
          "game_id": "wa30",
          "game_version": "ee6fef47",
          "published_at": "2026-06-19T01:34:28.847165Z",
          "score": 0,
          "model": "gpt-5-6-terra",
          "runner": "benchmark_agent",
          "config": "openai-gpt-5-6-terra-max",
          "verified_publish": true,
          "levels_completed": 0,
          "actions": 355,
          "resets": 62,
          "state": "NOT_FINISHED",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            0,
            0,
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            355,
            0,
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            71,
            119,
            183,
            98,
            368,
            68,
            79,
            442,
            415
          ]
        },
        {
          "session_id": "1d659839-4153-4347-873e-fe63eeb75015",
          "game_id": "wa30",
          "game_version": "ee6fef47",
          "published_at": "2026-06-13T22:54:52.702412Z",
          "score": 2.857166655310694,
          "model": "gpt-5-6-sol",
          "runner": "benchmark_agent",
          "config": "openai-gpt-5-6-sol-max",
          "verified_publish": true,
          "levels_completed": 2,
          "actions": 1009,
          "resets": 13,
          "state": "GAME_OVER",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            102.87755102040815,
            12.847474234286546,
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            70,
            332,
            607,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            71,
            119,
            183,
            98,
            368,
            68,
            79,
            442,
            415
          ]
        },
        {
          "session_id": "b73b1942-4de6-47e6-91a6-3b68031e452a",
          "game_id": "wa30",
          "game_version": "ee6fef47",
          "published_at": "2026-06-11T20:03:24.41586Z",
          "score": 2.2222222222222223,
          "model": "gpt-5-6-sol",
          "runner": "benchmark_agent",
          "config": "openai-gpt-5-6-sol-xhigh",
          "verified_publish": true,
          "levels_completed": 1,
          "actions": 661,
          "resets": 12,
          "state": "GAME_OVER",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            115,
            0,
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            66,
            595,
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            71,
            119,
            183,
            98,
            368,
            68,
            79,
            442,
            415
          ]
        },
        {
          "session_id": "c7f501aa-abfd-4a8c-b5d1-bf9d47e66c65",
          "game_id": "wa30",
          "game_version": "ee6fef47",
          "published_at": "2026-06-10T14:26:15.258472Z",
          "score": 1.8893948764078632,
          "model": "gpt-5-6-sol",
          "runner": "benchmark_agent",
          "config": "openai-gpt-5-6-sol-high",
          "verified_publish": true,
          "levels_completed": 1,
          "actions": 672,
          "resets": 10,
          "state": "GAME_OVER",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            85.02276943835385,
            0,
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            77,
            595,
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            71,
            119,
            183,
            98,
            368,
            68,
            79,
            442,
            415
          ]
        },
        {
          "session_id": "48f3b3fe-5d4e-4b6d-b4d0-ba6a6fe8073c",
          "game_id": "wa30",
          "game_version": "ee6fef47",
          "published_at": "2026-06-10T08:24:06.983316Z",
          "score": 0,
          "model": "gpt-5-6-sol",
          "runner": "benchmark_agent",
          "config": "openai-gpt-5-6-sol-medium",
          "verified_publish": true,
          "levels_completed": 0,
          "actions": 355,
          "resets": 1,
          "state": "GAME_OVER",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            0,
            0,
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            355,
            0,
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            71,
            119,
            183,
            98,
            368,
            68,
            79,
            442,
            415
          ]
        },
        {
          "session_id": "9498be77-a618-48da-935a-85ca2e9d0e32",
          "game_id": "wa30",
          "game_version": "ee6fef47",
          "published_at": "2026-06-09T22:04:20.551872Z",
          "score": 0,
          "model": "gpt-5-6-sol",
          "runner": "benchmark_agent",
          "config": "openai-gpt-5-6-sol-low",
          "verified_publish": true,
          "levels_completed": 0,
          "actions": 355,
          "resets": 2,
          "state": "GAME_OVER",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            0,
            0,
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            355,
            0,
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            71,
            119,
            183,
            98,
            368,
            68,
            79,
            442,
            415
          ]
        },
        {
          "session_id": "f394174b-3e6b-47b4-9878-2f0967ce1f6d",
          "game_id": "wa30",
          "game_version": "ee6fef47",
          "published_at": "2026-05-31T14:55:05.13282Z",
          "score": 1.8893948764078632,
          "model": "claude-opus-4-8",
          "runner": "benchmark_agent",
          "config": "anthropic-opus-4-8-high",
          "verified_publish": true,
          "levels_completed": 1,
          "actions": 672,
          "resets": 10,
          "state": "GAME_OVER",
          "completed": false,
          "number_of_levels": 9,
          "level_scores": [
            85.02276943835385,
            0,
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            77,
            595,
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            71,
            119,
            183,
            98,
            368,
            68,
            79,
            442,
            415
          ]
        },
        {
          "session_id": "2c1d140e-23fe-4b7c-bcc3-d5092fcd3660",
          "game_id": "wa30",
          "game_version": "ee6fef47",
          "published_at": "2026-05-01T05:09:18.774431Z",
          "score": 0,
          "model": "claude-opus-4-7",
          "runner": "benchmark_agent",
          "config": "anthropic-opus-4-7-high",
          "verified_publish": true,
          "levels_completed": 0,
          "actions": 355,
          "resets": 5,
          "state": "NOT_FINISHED",
          "completed": false,
          "number_of_levels": 9,
          "level_scores": [
            0,
            0,
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            355,
            0,
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            71,
            119,
            183,
            98,
            368,
            68,
            79,
            442,
            415
          ]
        },
        {
          "session_id": "04601c6e-3bd2-4515-b32c-ce9ae1c2e2eb",
          "game_id": "wa30",
          "game_version": "ee6fef47",
          "published_at": "2026-04-30T23:24:55.094395Z",
          "score": 0,
          "model": "gpt-5.5-2026-04-23",
          "runner": "benchmark_agent",
          "config": "openai-gpt-5-5-2026-04-23-high",
          "verified_publish": true,
          "levels_completed": 0,
          "actions": 355,
          "resets": 14,
          "state": "NOT_FINISHED",
          "completed": false,
          "number_of_levels": 9,
          "level_scores": [
            0,
            0,
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            355,
            0,
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            71,
            119,
            183,
            98,
            368,
            68,
            79,
            442,
            415
          ]
        },
        {
          "session_id": "86d9f75a-99be-487a-915a-0db0563d9c60",
          "game_id": "wa30",
          "game_version": "ee6fef47",
          "published_at": "2026-04-14T17:44:03.111389Z",
          "score": 0,
          "model": "grok-4.20-beta-0309-reasoning",
          "runner": "benchmark_agent",
          "config": "xai-grok-4-20-beta-0309-reasoning",
          "verified_publish": true,
          "levels_completed": 0,
          "actions": 322,
          "resets": 23,
          "state": "NOT_FINISHED",
          "completed": false,
          "number_of_levels": 9,
          "level_scores": [
            0,
            0,
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            322,
            0,
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            71,
            119,
            183,
            98,
            368,
            68,
            79,
            442,
            415
          ]
        },
        {
          "session_id": "ea30c6c0-b5e1-4d38-8117-749c8a0b86d2",
          "game_id": "wa30",
          "game_version": "ee6fef47",
          "published_at": "2026-04-14T17:44:02.763307Z",
          "score": 0,
          "model": "gpt-5.4-2026-03-05",
          "runner": "benchmark_agent",
          "config": "openai-gpt-5-4-2026-03-05-high",
          "verified_publish": true,
          "levels_completed": 0,
          "actions": 355,
          "resets": 1,
          "state": "GAME_OVER",
          "completed": false,
          "number_of_levels": 9,
          "level_scores": [
            0,
            0,
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            355,
            0,
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            71,
            119,
            183,
            98,
            368,
            68,
            79,
            442,
            415
          ]
        },
        {
          "session_id": "a4739589-d000-44ff-81df-49a6944aaabd",
          "game_id": "wa30",
          "game_version": "ee6fef47",
          "published_at": "2026-04-14T17:44:01.969803Z",
          "score": 0,
          "model": "gemini-3.1-pro-preview",
          "runner": "benchmark_agent",
          "config": "google-gemini-3-1-pro-preview",
          "verified_publish": true,
          "levels_completed": 0,
          "actions": 305,
          "resets": 3,
          "state": "GAME_OVER",
          "completed": false,
          "number_of_levels": 9,
          "level_scores": [
            0,
            0,
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            305,
            0,
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            71,
            119,
            183,
            98,
            368,
            68,
            79,
            442,
            415
          ]
        },
        {
          "session_id": "6f46c837-4f8e-41a9-ad72-743c3d905c3b",
          "game_id": "wa30",
          "game_version": "ee6fef47",
          "published_at": "2026-04-14T17:44:01.535634Z",
          "score": 0,
          "model": "claude-opus-4-6",
          "runner": "benchmark_agent",
          "config": "anthropic-opus-4-6-max-effort",
          "verified_publish": true,
          "levels_completed": 0,
          "actions": 355,
          "resets": 9,
          "state": "NOT_FINISHED",
          "completed": false,
          "number_of_levels": 9,
          "level_scores": [
            0,
            0,
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            355,
            0,
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            71,
            119,
            183,
            98,
            368,
            68,
            79,
            442,
            415
          ]
        }
      ]
    },
    {
      "game": "re86",
      "source": "https://arcprize.org/tasks/re86",
      "runs": [
        {
          "session_id": "e9babb6e-6395-412d-a33a-f742fd76b6e5",
          "game_id": "re86",
          "game_version": "8af5384d",
          "published_at": "2026-09-25T15:25:22.654905Z",
          "score": 16.666666666666664,
          "model": "gpt-6-sol",
          "runner": "benchmark_agent",
          "config": "openai-gpt-6-sol-max",
          "verified_publish": true,
          "levels_completed": 3,
          "actions": 365,
          "resets": 1,
          "state": "GAME_OVER",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            92.72976680384086,
            100,
            115,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            27,
            42,
            73,
            223,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            26,
            42,
            86,
            108,
            189,
            139,
            424,
            241
          ]
        },
        {
          "session_id": "64f94388-c230-43b4-9c68-62b68866072e",
          "game_id": "re86",
          "game_version": "8af5384d",
          "published_at": "2026-09-25T09:02:18.074595Z",
          "score": 8.333333333333332,
          "model": "gpt-6-sol",
          "runner": "benchmark_agent",
          "config": "openai-gpt-6-sol-xhigh",
          "verified_publish": true,
          "levels_completed": 2,
          "actions": 496,
          "resets": 2,
          "state": "GAME_OVER",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            108.16000000000001,
            104.93753718024985,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            25,
            41,
            430,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            26,
            42,
            86,
            108,
            189,
            139,
            424,
            241
          ]
        },
        {
          "session_id": "8c0d2e3f-683a-4309-aed0-76a9a2263dca",
          "game_id": "re86",
          "game_version": "8af5384d",
          "published_at": "2026-09-25T07:55:55.731471Z",
          "score": 14.985712470302651,
          "model": "gpt-6-sol",
          "runner": "benchmark_agent",
          "config": "openai-gpt-6-sol-high-provider-adapter",
          "verified_publish": true,
          "levels_completed": 4,
          "actions": 1380,
          "resets": 5,
          "state": "GAME_OVER",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            108.16000000000001,
            7.635041551246539,
            43.76331360946745,
            71.19140625,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            25,
            152,
            130,
            128,
            945,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            26,
            42,
            86,
            108,
            189,
            139,
            424,
            241
          ]
        },
        {
          "session_id": "c098f36a-79f7-4649-9cfd-fafe1961ae30",
          "game_id": "re86",
          "game_version": "8af5384d",
          "published_at": "2026-09-25T07:43:04.365534Z",
          "score": 33.61307585253784,
          "model": "gpt-6-sol",
          "runner": "benchmark_agent",
          "config": "openai-gpt-6-sol-xhigh-provider-adapter",
          "verified_publish": true,
          "levels_completed": 5,
          "actions": 1332,
          "resets": 4,
          "state": "GAME_OVER",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            108.16000000000001,
            104.93753718024985,
            115,
            115,
            17.407131266172538,
            0,
            0,
            0
          ],
          "level_actions": [
            25,
            41,
            48,
            70,
            453,
            695,
            0,
            0
          ],
          "level_baseline_actions": [
            26,
            42,
            86,
            108,
            189,
            139,
            424,
            241
          ]
        },
        {
          "session_id": "98160d3a-4d6f-47ae-99ed-16416ca954db",
          "game_id": "re86",
          "game_version": "8af5384d",
          "published_at": "2026-09-25T06:33:55.564309Z",
          "score": 0,
          "model": "gpt-6-sol",
          "runner": "benchmark_agent",
          "config": "openai-gpt-6-sol-none-provider-adapter",
          "verified_publish": true,
          "levels_completed": 0,
          "actions": 130,
          "resets": 1,
          "state": "GAME_OVER",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            0,
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            130,
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            26,
            42,
            86,
            108,
            189,
            139,
            424,
            241
          ]
        },
        {
          "session_id": "9febaef9-861f-440c-b1b8-7a55d6974b2e",
          "game_id": "re86",
          "game_version": "8af5384d",
          "published_at": "2026-09-25T06:31:53.534769Z",
          "score": 4.570128432920168,
          "model": "gpt-6-sol",
          "runner": "benchmark_agent",
          "config": "openai-gpt-6-sol-medium-provider-adapter",
          "verified_publish": true,
          "levels_completed": 2,
          "actions": 541,
          "resets": 2,
          "state": "GAME_OVER",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            17.58584807492196,
            73.46938775510203,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            62,
            49,
            430,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            26,
            42,
            86,
            108,
            189,
            139,
            424,
            241
          ]
        },
        {
          "session_id": "34d9512b-14c7-4aef-a9c2-e28fecb899b3",
          "game_id": "re86",
          "game_version": "8af5384d",
          "published_at": "2026-09-25T06:23:20.875926Z",
          "score": 0,
          "model": "gpt-6-sol",
          "runner": "benchmark_agent",
          "config": "openai-gpt-6-sol-none",
          "verified_publish": true,
          "levels_completed": 0,
          "actions": 130,
          "resets": 1,
          "state": "GAME_OVER",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            0,
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            130,
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            26,
            42,
            86,
            108,
            189,
            139,
            424,
            241
          ]
        },
        {
          "session_id": "0dca5a59-8edb-4376-af09-408b8aedf154",
          "game_id": "re86",
          "game_version": "8af5384d",
          "published_at": "2026-09-25T06:17:09.550522Z",
          "score": 2.232791650151936,
          "model": "gpt-6-sol",
          "runner": "benchmark_agent",
          "config": "openai-gpt-6-sol-low-provider-adapter",
          "verified_publish": true,
          "levels_completed": 1,
          "actions": 239,
          "resets": 2,
          "state": "GAME_OVER",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            80.38049940546969,
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            29,
            210,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            26,
            42,
            86,
            108,
            189,
            139,
            424,
            241
          ]
        },
        {
          "session_id": "5481ed3b-2a4c-4721-9d1e-a6a308b7d3b9",
          "game_id": "re86",
          "game_version": "8af5384d",
          "published_at": "2026-09-25T05:58:02.578426Z",
          "score": 0.15804879873560962,
          "model": "gpt-6-sol",
          "runner": "benchmark_agent",
          "config": "openai-gpt-6-sol-medium",
          "verified_publish": true,
          "levels_completed": 1,
          "actions": 319,
          "resets": 3,
          "state": "GAME_OVER",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            5.689756754481946,
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            109,
            210,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            26,
            42,
            86,
            108,
            189,
            139,
            424,
            241
          ]
        },
        {
          "session_id": "b451636d-1547-4c8f-a986-1c73e7f80b68",
          "game_id": "re86",
          "game_version": "8af5384d",
          "published_at": "2026-09-25T05:56:30.438521Z",
          "score": 0,
          "model": "gpt-6-sol",
          "runner": "benchmark_agent",
          "config": "openai-gpt-6-sol-low",
          "verified_publish": true,
          "levels_completed": 0,
          "actions": 130,
          "resets": 1,
          "state": "GAME_OVER",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            0,
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            130,
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            26,
            42,
            86,
            108,
            189,
            139,
            424,
            241
          ]
        },
        {
          "session_id": "1f37cbb7-8966-4457-95fe-8b078e823e0c",
          "game_id": "re86",
          "game_version": "8af5384d",
          "published_at": "2026-09-25T05:43:04.545821Z",
          "score": 2.7777777777777777,
          "model": "gpt-6-sol",
          "runner": "benchmark_agent",
          "config": "openai-gpt-6-sol-high",
          "verified_publish": true,
          "levels_completed": 1,
          "actions": 235,
          "resets": 3,
          "state": "GAME_OVER",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            108.16000000000001,
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            25,
            210,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            26,
            42,
            86,
            108,
            189,
            139,
            424,
            241
          ]
        },
        {
          "session_id": "427ac58e-a22f-4b30-b35a-ed2c9d666931",
          "game_id": "re86",
          "game_version": "8af5384d",
          "published_at": "2026-09-21T16:59:07.645768Z",
          "score": 67.89728274175039,
          "model": "gemini-3.8-flash",
          "runner": "benchmark_agent",
          "config": "google-gemini-3-8-flash-low-provider-adapter",
          "verified_publish": true,
          "levels_completed": 8,
          "actions": 2212,
          "resets": 9,
          "state": "WIN",
          "completed": true,
          "number_of_levels": 0,
          "level_scores": [
            115,
            115,
            115,
            115,
            115,
            51.86984885500282,
            53.625898980136675,
            4.0877240890049995
          ],
          "level_actions": [
            24,
            38,
            53,
            54,
            79,
            193,
            579,
            1192
          ],
          "level_baseline_actions": [
            26,
            42,
            86,
            108,
            189,
            139,
            424,
            241
          ]
        },
        {
          "session_id": "5b915336-4ea6-428b-a2e7-abd148eb9ecc",
          "game_id": "re86",
          "game_version": "8af5384d",
          "published_at": "2026-09-20T15:11:18.353762Z",
          "score": 58.333333333333336,
          "model": "gemini-3.8-flash",
          "runner": "benchmark_agent",
          "config": "google-gemini-3-8-flash-medium-provider-adapter",
          "verified_publish": true,
          "levels_completed": 6,
          "actions": 2465,
          "resets": 14,
          "state": "GAME_OVER",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            115,
            115,
            115,
            115,
            115,
            115,
            0,
            0
          ],
          "level_actions": [
            21,
            36,
            53,
            81,
            76,
            78,
            2120,
            0
          ],
          "level_baseline_actions": [
            26,
            42,
            86,
            108,
            189,
            139,
            424,
            241
          ]
        },
        {
          "session_id": "3b748248-a58c-40a9-99aa-1562339956d8",
          "game_id": "re86",
          "game_version": "8af5384d",
          "published_at": "2026-09-17T22:37:46.229742Z",
          "score": 100,
          "model": "gemini-3.8-flash",
          "runner": "benchmark_agent",
          "config": "google-gemini-3-8-flash-high-provider-adapter",
          "verified_publish": true,
          "levels_completed": 8,
          "actions": 840,
          "resets": 0,
          "state": "WIN",
          "completed": true,
          "number_of_levels": 0,
          "level_scores": [
            62.0752984389348,
            115,
            115,
            115,
            115,
            94.48383784048119,
            115,
            73.03581308787284
          ],
          "level_actions": [
            33,
            36,
            49,
            76,
            66,
            143,
            155,
            282
          ],
          "level_baseline_actions": [
            26,
            42,
            86,
            108,
            189,
            139,
            424,
            241
          ]
        },
        {
          "session_id": "c8855620-e353-4cbc-8f08-e99181caf6a0",
          "game_id": "re86",
          "game_version": "8af5384d",
          "published_at": "2026-09-16T19:25:13.662153Z",
          "score": 0,
          "model": "gpt-6-luna",
          "runner": "benchmark_agent",
          "config": "openai-gpt-6-luna-low",
          "verified_publish": true,
          "levels_completed": 0,
          "actions": 130,
          "resets": 1,
          "state": "GAME_OVER",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            0,
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            130,
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            26,
            42,
            86,
            108,
            189,
            139,
            424,
            241
          ]
        },
        {
          "session_id": "59230a03-02f1-46e3-9551-ad9a7a458a93",
          "game_id": "re86",
          "game_version": "8af5384d",
          "published_at": "2026-09-16T19:24:55.513985Z",
          "score": 0,
          "model": "gpt-6-luna",
          "runner": "benchmark_agent",
          "config": "openai-gpt-6-luna-none-provider-adapter",
          "verified_publish": true,
          "levels_completed": 0,
          "actions": 130,
          "resets": 2,
          "state": "GAME_OVER",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            0,
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            130,
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            26,
            42,
            86,
            108,
            189,
            139,
            424,
            241
          ]
        },
        {
          "session_id": "ae4dcf99-7463-437c-b0da-f9a29ea20056",
          "game_id": "re86",
          "game_version": "8af5384d",
          "published_at": "2026-09-16T19:23:59.735464Z",
          "score": 0,
          "model": "gpt-6-luna",
          "runner": "benchmark_agent",
          "config": "openai-gpt-6-luna-low-provider-adapter",
          "verified_publish": true,
          "levels_completed": 0,
          "actions": 130,
          "resets": 1,
          "state": "GAME_OVER",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            0,
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            130,
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            26,
            42,
            86,
            108,
            189,
            139,
            424,
            241
          ]
        },
        {
          "session_id": "63f4d272-0391-4c53-9501-76fb00599ac3",
          "game_id": "re86",
          "game_version": "8af5384d",
          "published_at": "2026-09-16T19:23:16.730961Z",
          "score": 0,
          "model": "gpt-6-luna",
          "runner": "benchmark_agent",
          "config": "openai-gpt-6-luna-none",
          "verified_publish": true,
          "levels_completed": 0,
          "actions": 130,
          "resets": 6,
          "state": "NOT_FINISHED",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            0,
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            130,
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            26,
            42,
            86,
            108,
            189,
            139,
            424,
            241
          ]
        },
        {
          "session_id": "847acaaa-39f8-4e1e-a19f-9dcf3a187a01",
          "game_id": "re86",
          "game_version": "8af5384d",
          "published_at": "2026-09-16T14:47:31.916599Z",
          "score": 77.77777777777779,
          "model": "gpt-6-1-sol",
          "runner": "benchmark_agent",
          "config": "openai-gpt-6-1-sol-xhigh",
          "verified_publish": true,
          "levels_completed": 7,
          "actions": 1811,
          "resets": 9,
          "state": "NOT_FINISHED",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            115,
            115,
            115,
            115,
            115,
            115,
            115,
            0
          ],
          "level_actions": [
            20,
            36,
            47,
            44,
            65,
            65,
            388,
            1146
          ],
          "level_baseline_actions": [
            26,
            42,
            86,
            108,
            189,
            139,
            424,
            241
          ]
        },
        {
          "session_id": "5e3b8c1c-a615-4e43-816f-ffda31e86edc",
          "game_id": "re86",
          "game_version": "8af5384d",
          "published_at": "2026-09-16T09:54:58.281251Z",
          "score": 35.20543499683615,
          "model": "gpt-6-1-sol",
          "runner": "benchmark_agent",
          "config": "openai-gpt-6-1-sol-low",
          "verified_publish": true,
          "levels_completed": 6,
          "actions": 1981,
          "resets": 5,
          "state": "GAME_OVER",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            115,
            115,
            115,
            115,
            7.590587255310285,
            13.24045393492503,
            0,
            0
          ],
          "level_actions": [
            21,
            39,
            59,
            66,
            686,
            382,
            728,
            0
          ],
          "level_baseline_actions": [
            26,
            42,
            86,
            108,
            189,
            139,
            424,
            241
          ]
        },
        {
          "session_id": "01722343-77e7-4be3-94b7-f49702273950",
          "game_id": "re86",
          "game_version": "8af5384d",
          "published_at": "2026-09-16T09:08:47.038209Z",
          "score": 58.333333333333336,
          "model": "gpt-6-1-sol",
          "runner": "benchmark_agent",
          "config": "openai-gpt-6-1-sol-medium",
          "verified_publish": true,
          "levels_completed": 6,
          "actions": 2543,
          "resets": 10,
          "state": "GAME_OVER",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            115,
            115,
            115,
            115,
            95.89787645305915,
            115,
            0,
            0
          ],
          "level_actions": [
            20,
            36,
            60,
            71,
            193,
            113,
            2050,
            0
          ],
          "level_baseline_actions": [
            26,
            42,
            86,
            108,
            189,
            139,
            424,
            241
          ]
        },
        {
          "session_id": "743e2d2c-eefc-4e64-9012-03608f6ddc3a",
          "game_id": "re86",
          "game_version": "8af5384d",
          "published_at": "2026-09-16T09:08:39.545051Z",
          "score": 69.83065210223049,
          "model": "gpt-6-1-sol",
          "runner": "benchmark_agent",
          "config": "openai-gpt-6-1-sol-high",
          "verified_publish": true,
          "levels_completed": 7,
          "actions": 1605,
          "resets": 13,
          "state": "NOT_FINISHED",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            115,
            115,
            115,
            115,
            115,
            115,
            14.129067954328253,
            0
          ],
          "level_actions": [
            20,
            36,
            47,
            45,
            69,
            88,
            1128,
            172
          ],
          "level_baseline_actions": [
            26,
            42,
            86,
            108,
            189,
            139,
            424,
            241
          ]
        },
        {
          "session_id": "83112511-c7f1-48b6-935e-62918f351dac",
          "game_id": "re86",
          "game_version": "8af5384d",
          "published_at": "2026-09-16T09:01:32.669351Z",
          "score": 0,
          "model": "gpt-6-luna",
          "runner": "benchmark_agent",
          "config": "openai-gpt-6-luna-medium-provider-adapter",
          "verified_publish": true,
          "levels_completed": 0,
          "actions": 130,
          "resets": 1,
          "state": "GAME_OVER",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            0,
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            130,
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            26,
            42,
            86,
            108,
            189,
            139,
            424,
            241
          ]
        },
        {
          "session_id": "4f88c563-f361-45c6-ab26-64a74b3faa5f",
          "game_id": "re86",
          "game_version": "8af5384d",
          "published_at": "2026-09-16T08:50:55.00267Z",
          "score": 0,
          "model": "gpt-6-luna",
          "runner": "benchmark_agent",
          "config": "openai-gpt-6-luna-medium",
          "verified_publish": true,
          "levels_completed": 0,
          "actions": 130,
          "resets": 1,
          "state": "GAME_OVER",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            0,
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            130,
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            26,
            42,
            86,
            108,
            189,
            139,
            424,
            241
          ]
        },
        {
          "session_id": "82ee5f9c-770c-4aea-bab2-33021cff6e2c",
          "game_id": "re86",
          "game_version": "8af5384d",
          "published_at": "2026-09-16T08:31:32.198572Z",
          "score": 0,
          "model": "gpt-6-luna",
          "runner": "benchmark_agent",
          "config": "openai-gpt-6-luna-high",
          "verified_publish": true,
          "levels_completed": 0,
          "actions": 130,
          "resets": 3,
          "state": "GAME_OVER",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            0,
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            130,
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            26,
            42,
            86,
            108,
            189,
            139,
            424,
            241
          ]
        },
        {
          "session_id": "5e51e0b6-db2f-4c3e-ad48-94af775490c0",
          "game_id": "re86",
          "game_version": "8af5384d",
          "published_at": "2026-09-16T08:27:38.769538Z",
          "score": 0,
          "model": "gpt-6-luna",
          "runner": "benchmark_agent",
          "config": "openai-gpt-6-luna-high-provider-adapter",
          "verified_publish": true,
          "levels_completed": 0,
          "actions": 130,
          "resets": 1,
          "state": "NOT_FINISHED",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            0,
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            130,
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            26,
            42,
            86,
            108,
            189,
            139,
            424,
            241
          ]
        },
        {
          "session_id": "42310efb-c191-495c-ba7f-ea1d04f06cd0",
          "game_id": "re86",
          "game_version": "8af5384d",
          "published_at": "2026-09-16T07:40:56.433579Z",
          "score": 0,
          "model": "gpt-6-luna",
          "runner": "benchmark_agent",
          "config": "openai-gpt-6-luna-xhigh-provider-adapter",
          "verified_publish": true,
          "levels_completed": 0,
          "actions": 130,
          "resets": 1,
          "state": "GAME_OVER",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            0,
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            130,
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            26,
            42,
            86,
            108,
            189,
            139,
            424,
            241
          ]
        },
        {
          "session_id": "2ef09b85-3ce3-46d3-81bd-e149b4259232",
          "game_id": "re86",
          "game_version": "8af5384d",
          "published_at": "2026-09-16T07:24:56.32544Z",
          "score": 77.77777777777779,
          "model": "gpt-6-1-sol",
          "runner": "benchmark_agent",
          "config": "openai-gpt-6-1-sol-max",
          "verified_publish": true,
          "levels_completed": 7,
          "actions": 871,
          "resets": 3,
          "state": "NOT_FINISHED",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            115,
            115,
            115,
            115,
            115,
            115,
            115,
            0
          ],
          "level_actions": [
            20,
            36,
            47,
            44,
            63,
            60,
            255,
            346
          ],
          "level_baseline_actions": [
            26,
            42,
            86,
            108,
            189,
            139,
            424,
            241
          ]
        },
        {
          "session_id": "9bbed442-6a96-4a44-b614-fd8bba450d3b",
          "game_id": "re86",
          "game_version": "8af5384d",
          "published_at": "2026-09-16T07:24:43.487435Z",
          "score": 0,
          "model": "gpt-6-luna",
          "runner": "benchmark_agent",
          "config": "openai-gpt-6-luna-xhigh",
          "verified_publish": true,
          "levels_completed": 0,
          "actions": 130,
          "resets": 1,
          "state": "NOT_FINISHED",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            0,
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            130,
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            26,
            42,
            86,
            108,
            189,
            139,
            424,
            241
          ]
        },
        {
          "session_id": "f1607529-dea1-46fa-8947-5c9b826fe7b3",
          "game_id": "re86",
          "game_version": "8af5384d",
          "published_at": "2026-09-16T07:24:19.751212Z",
          "score": 0,
          "model": "gpt-6-luna",
          "runner": "benchmark_agent",
          "config": "openai-gpt-6-luna-max",
          "verified_publish": true,
          "levels_completed": 0,
          "actions": 130,
          "resets": 2,
          "state": "NOT_FINISHED",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            0,
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            130,
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            26,
            42,
            86,
            108,
            189,
            139,
            424,
            241
          ]
        },
        {
          "session_id": "07f9cc11-e9f4-4c3f-879d-adfaf0ca40d2",
          "game_id": "re86",
          "game_version": "8af5384d",
          "published_at": "2026-09-16T05:17:53.16975Z",
          "score": 100,
          "model": "gpt-6-1-sol",
          "runner": "benchmark_agent",
          "config": "openai-gpt-6-1-sol-high-provider-adapter",
          "verified_publish": true,
          "levels_completed": 8,
          "actions": 584,
          "resets": 0,
          "state": "WIN",
          "completed": true,
          "number_of_levels": 0,
          "level_scores": [
            115,
            115,
            115,
            115,
            115,
            115,
            115,
            115
          ],
          "level_actions": [
            20,
            36,
            47,
            44,
            63,
            69,
            104,
            201
          ],
          "level_baseline_actions": [
            26,
            42,
            86,
            108,
            189,
            139,
            424,
            241
          ]
        },
        {
          "session_id": "c73d055e-e31f-4b15-b3de-aa5f5078277b",
          "game_id": "re86",
          "game_version": "8af5384d",
          "published_at": "2026-09-16T04:11:51.339215Z",
          "score": 100,
          "model": "gpt-6-1-sol",
          "runner": "benchmark_agent",
          "config": "openai-gpt-6-1-sol-xhigh-provider-adapter",
          "verified_publish": true,
          "levels_completed": 8,
          "actions": 614,
          "resets": 0,
          "state": "WIN",
          "completed": true,
          "number_of_levels": 0,
          "level_scores": [
            115,
            115,
            115,
            115,
            115,
            115,
            115,
            115
          ],
          "level_actions": [
            20,
            36,
            47,
            45,
            63,
            63,
            130,
            210
          ],
          "level_baseline_actions": [
            26,
            42,
            86,
            108,
            189,
            139,
            424,
            241
          ]
        },
        {
          "session_id": "0284558d-fce1-4ee7-a6e3-205b08c1027e",
          "game_id": "re86",
          "game_version": "8af5384d",
          "published_at": "2026-09-16T02:53:24.275193Z",
          "score": 0.5779556102732464,
          "model": "gpt-6-luna",
          "runner": "benchmark_agent",
          "config": "openai-gpt-6-luna-max-provider-adapter",
          "verified_publish": true,
          "levels_completed": 1,
          "actions": 267,
          "resets": 5,
          "state": "GAME_OVER",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            20.80640196983687,
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            57,
            210,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            26,
            42,
            86,
            108,
            189,
            139,
            424,
            241
          ]
        },
        {
          "session_id": "c137ebfc-caa3-4c5e-a5a6-04edca56b47f",
          "game_id": "re86",
          "game_version": "8af5384d",
          "published_at": "2026-09-16T02:17:20.976798Z",
          "score": 100,
          "model": "gpt-6-1-sol",
          "runner": "benchmark_agent",
          "config": "openai-gpt-6-1-sol-max-provider-adapter",
          "verified_publish": true,
          "levels_completed": 8,
          "actions": 614,
          "resets": 0,
          "state": "WIN",
          "completed": true,
          "number_of_levels": 0,
          "level_scores": [
            115,
            115,
            115,
            115,
            115,
            115,
            115,
            115
          ],
          "level_actions": [
            20,
            36,
            47,
            44,
            63,
            79,
            125,
            200
          ],
          "level_baseline_actions": [
            26,
            42,
            86,
            108,
            189,
            139,
            424,
            241
          ]
        },
        {
          "session_id": "69cf72d3-8c76-44d9-be86-b59e44b5114b",
          "game_id": "re86",
          "game_version": "8af5384d",
          "published_at": "2026-09-12T11:02:06.134997Z",
          "score": 27.77777777777778,
          "model": "gemini-3.8-flash",
          "runner": "benchmark_agent",
          "config": "google-gemini-3-8-flash-high",
          "verified_publish": true,
          "levels_completed": 4,
          "actions": 264,
          "resets": 0,
          "state": "NOT_FINISHED",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            108.16000000000001,
            87.11111111111111,
            115,
            115,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            25,
            45,
            77,
            94,
            23,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            26,
            42,
            86,
            108,
            189,
            139,
            424,
            241
          ]
        },
        {
          "session_id": "2b75a18c-ea41-440c-9a39-6f680ecfeb27",
          "game_id": "re86",
          "game_version": "8af5384d",
          "published_at": "2026-09-11T05:57:21.410218Z",
          "score": 8.333333333333332,
          "model": "gemini-3.8-flash",
          "runner": "benchmark_agent",
          "config": "google-gemini-3-8-flash-low",
          "verified_publish": true,
          "levels_completed": 2,
          "actions": 85,
          "resets": 0,
          "state": "NOT_FINISHED",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            92.72976680384086,
            115,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            27,
            36,
            22,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            26,
            42,
            86,
            108,
            189,
            139,
            424,
            241
          ]
        },
        {
          "session_id": "47758e6e-2561-4d65-b249-83fc9f76d0db",
          "game_id": "re86",
          "game_version": "8af5384d",
          "published_at": "2026-09-10T23:32:27.553246Z",
          "score": 8.333333333333332,
          "model": "gemini-3.8-flash",
          "runner": "benchmark_agent",
          "config": "google-gemini-3-8-flash-medium",
          "verified_publish": true,
          "levels_completed": 2,
          "actions": 493,
          "resets": 4,
          "state": "NOT_FINISHED",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            115,
            110.25,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            23,
            40,
            430,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            26,
            42,
            86,
            108,
            189,
            139,
            424,
            241
          ]
        },
        {
          "session_id": "1c9fea8c-a8de-4406-9e68-27d7eb737377",
          "game_id": "re86",
          "game_version": "8af5384d",
          "published_at": "2026-08-11T18:10:23.696954Z",
          "score": 27.77777777777778,
          "model": "grok-4.6",
          "runner": "benchmark_agent",
          "config": "xai-grok-4-6-xhigh",
          "verified_publish": true,
          "levels_completed": 4,
          "actions": 955,
          "resets": 7,
          "state": "GAME_OVER",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            62.0752984389348,
            115,
            115,
            115,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            33,
            39,
            49,
            50,
            784,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            26,
            42,
            86,
            108,
            189,
            139,
            424,
            241
          ]
        },
        {
          "session_id": "29b2313c-2135-489d-a839-dcc01cd7dc21",
          "game_id": "re86",
          "game_version": "8af5384d",
          "published_at": "2026-07-21T12:10:55.358816Z",
          "score": 58.333333333333336,
          "model": "claude-opus-5",
          "runner": "benchmark_agent",
          "config": "anthropic-claude-opus-5-high",
          "verified_publish": true,
          "levels_completed": 6,
          "actions": 1112,
          "resets": 3,
          "state": "GAME_OVER",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            115,
            115,
            115,
            82.36706447284796,
            115,
            115,
            0,
            0
          ],
          "level_actions": [
            20,
            36,
            47,
            119,
            63,
            84,
            743,
            0
          ],
          "level_baseline_actions": [
            26,
            42,
            86,
            108,
            189,
            139,
            424,
            241
          ]
        },
        {
          "session_id": "2409ee3c-df20-4e4d-8a5b-911195fde490",
          "game_id": "re86",
          "game_version": "8af5384d",
          "published_at": "2026-06-24T10:50:02.7055Z",
          "score": 0,
          "model": "gpt-5-6-luna",
          "runner": "benchmark_agent",
          "config": "openai-gpt-5-6-luna-low",
          "verified_publish": true,
          "levels_completed": 0,
          "actions": 130,
          "resets": 2,
          "state": "GAME_OVER",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            0,
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            130,
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            26,
            42,
            86,
            108,
            189,
            139,
            424,
            241
          ]
        },
        {
          "session_id": "9ed05fc0-1708-4eec-b117-02f5ae0c4fb7",
          "game_id": "re86",
          "game_version": "8af5384d",
          "published_at": "2026-06-24T10:22:57.000938Z",
          "score": 0,
          "model": "gpt-5-6-luna",
          "runner": "benchmark_agent",
          "config": "openai-gpt-5-6-luna-medium",
          "verified_publish": true,
          "levels_completed": 0,
          "actions": 101,
          "resets": 1,
          "state": "GAME_OVER",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            0,
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            101,
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            26,
            42,
            86,
            108,
            189,
            139,
            424,
            241
          ]
        },
        {
          "session_id": "8b0fe410-44cb-4e28-915e-9d93e63af544",
          "game_id": "re86",
          "game_version": "8af5384d",
          "published_at": "2026-06-24T03:49:44.354008Z",
          "score": 0,
          "model": "gpt-5-6-luna",
          "runner": "benchmark_agent",
          "config": "openai-gpt-5-6-luna-high",
          "verified_publish": true,
          "levels_completed": 0,
          "actions": 130,
          "resets": 8,
          "state": "NOT_FINISHED",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            0,
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            130,
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            26,
            42,
            86,
            108,
            189,
            139,
            424,
            241
          ]
        },
        {
          "session_id": "c7db0424-e99d-442a-941d-421eb3a58bef",
          "game_id": "re86",
          "game_version": "8af5384d",
          "published_at": "2026-06-24T00:52:28.67094Z",
          "score": 0,
          "model": "gpt-5-6-luna",
          "runner": "benchmark_agent",
          "config": "openai-gpt-5-6-luna-xhigh",
          "verified_publish": true,
          "levels_completed": 0,
          "actions": 130,
          "resets": 14,
          "state": "NOT_FINISHED",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            0,
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            130,
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            26,
            42,
            86,
            108,
            189,
            139,
            424,
            241
          ]
        },
        {
          "session_id": "b711fc21-a840-4d92-a8a1-3baed64b65c7",
          "game_id": "re86",
          "game_version": "8af5384d",
          "published_at": "2026-06-23T02:48:21.00527Z",
          "score": 0,
          "model": "gpt-5-6-luna",
          "runner": "benchmark_agent",
          "config": "openai-gpt-5-6-luna-max",
          "verified_publish": true,
          "levels_completed": 0,
          "actions": 130,
          "resets": 22,
          "state": "NOT_FINISHED",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            0,
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            130,
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            26,
            42,
            86,
            108,
            189,
            139,
            424,
            241
          ]
        },
        {
          "session_id": "2730723a-c1d7-4817-80b5-d5377516b99f",
          "game_id": "re86",
          "game_version": "8af5384d",
          "published_at": "2026-06-20T11:58:24.079764Z",
          "score": 0,
          "model": "gpt-5-6-terra",
          "runner": "benchmark_agent",
          "config": "openai-gpt-5-6-terra-low",
          "verified_publish": true,
          "levels_completed": 0,
          "actions": 130,
          "resets": 3,
          "state": "GAME_OVER",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            0,
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            130,
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            26,
            42,
            86,
            108,
            189,
            139,
            424,
            241
          ]
        },
        {
          "session_id": "1072ae6c-e75e-489a-80a8-84b11483ea14",
          "game_id": "re86",
          "game_version": "8af5384d",
          "published_at": "2026-06-20T09:24:44.418921Z",
          "score": 0.504643315715608,
          "model": "gpt-5-6-terra",
          "runner": "benchmark_agent",
          "config": "openai-gpt-5-6-terra-medium",
          "verified_publish": true,
          "levels_completed": 1,
          "actions": 271,
          "resets": 2,
          "state": "GAME_OVER",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            18.16715936576189,
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            61,
            210,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            26,
            42,
            86,
            108,
            189,
            139,
            424,
            241
          ]
        },
        {
          "session_id": "dee474b6-aed3-4b08-97d4-a5be0921dfbb",
          "game_id": "re86",
          "game_version": "8af5384d",
          "published_at": "2026-06-20T08:29:17.869829Z",
          "score": 0,
          "model": "gpt-5-6-terra",
          "runner": "benchmark_agent",
          "config": "openai-gpt-5-6-terra-high",
          "verified_publish": true,
          "levels_completed": 0,
          "actions": 130,
          "resets": 1,
          "state": "GAME_OVER",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            0,
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            130,
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            26,
            42,
            86,
            108,
            189,
            139,
            424,
            241
          ]
        },
        {
          "session_id": "1e9ccd7b-8646-4ceb-96dc-e3f4e9c7cfc7",
          "game_id": "re86",
          "game_version": "8af5384d",
          "published_at": "2026-06-20T05:13:15.504924Z",
          "score": 0.4060938100730488,
          "model": "gpt-5-6-terra",
          "runner": "benchmark_agent",
          "config": "openai-gpt-5-6-terra-xhigh",
          "verified_publish": true,
          "levels_completed": 1,
          "actions": 278,
          "resets": 13,
          "state": "NOT_FINISHED",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            14.619377162629757,
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            68,
            210,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            26,
            42,
            86,
            108,
            189,
            139,
            424,
            241
          ]
        },
        {
          "session_id": "d728a33e-f8c8-45e1-b8f9-6450aa916e5e",
          "game_id": "re86",
          "game_version": "8af5384d",
          "published_at": "2026-06-19T09:02:31.30392Z",
          "score": 0.9322512904654165,
          "model": "gpt-5-6-terra",
          "runner": "benchmark_agent",
          "config": "openai-gpt-5-6-terra-max",
          "verified_publish": true,
          "levels_completed": 2,
          "actions": 633,
          "resets": 59,
          "state": "NOT_FINISHED",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            12.685306811784574,
            10.437869822485208,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            73,
            130,
            430,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            26,
            42,
            86,
            108,
            189,
            139,
            424,
            241
          ]
        },
        {
          "session_id": "09cb85c4-e9e5-44eb-8326-4429aa2cb7d4",
          "game_id": "re86",
          "game_version": "8af5384d",
          "published_at": "2026-06-13T18:03:52.40854Z",
          "score": 16.666666666666664,
          "model": "gpt-5-6-sol",
          "runner": "benchmark_agent",
          "config": "openai-gpt-5-6-sol-max",
          "verified_publish": true,
          "levels_completed": 3,
          "actions": 668,
          "resets": 6,
          "state": "NOT_FINISHED",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            108.16000000000001,
            91.11570247933885,
            115,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            25,
            44,
            59,
            540,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            26,
            42,
            86,
            108,
            189,
            139,
            424,
            241
          ]
        },
        {
          "session_id": "3723920e-cbf4-4e90-a1e3-22145c5f8c64",
          "game_id": "re86",
          "game_version": "8af5384d",
          "published_at": "2026-06-11T19:57:47.736029Z",
          "score": 8.333333333333332,
          "model": "gpt-5-6-sol",
          "runner": "benchmark_agent",
          "config": "openai-gpt-5-6-sol-xhigh",
          "verified_publish": true,
          "levels_completed": 2,
          "actions": 492,
          "resets": 6,
          "state": "NOT_FINISHED",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            108.16000000000001,
            115,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            25,
            37,
            430,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            26,
            42,
            86,
            108,
            189,
            139,
            424,
            241
          ]
        },
        {
          "session_id": "15ad2916-17e6-4a43-9840-6d6ab404d51d",
          "game_id": "re86",
          "game_version": "8af5384d",
          "published_at": "2026-06-10T11:24:44.240863Z",
          "score": 2.7777777777777777,
          "model": "gpt-5-6-sol",
          "runner": "benchmark_agent",
          "config": "openai-gpt-5-6-sol-high",
          "verified_publish": true,
          "levels_completed": 1,
          "actions": 236,
          "resets": 2,
          "state": "GAME_OVER",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            100,
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            26,
            210,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            26,
            42,
            86,
            108,
            189,
            139,
            424,
            241
          ]
        },
        {
          "session_id": "ca4e0543-4429-418d-b8f8-aa89ab2f47f2",
          "game_id": "re86",
          "game_version": "8af5384d",
          "published_at": "2026-06-10T08:17:57.247228Z",
          "score": 2.3951247165532883,
          "model": "gpt-5-6-sol",
          "runner": "benchmark_agent",
          "config": "openai-gpt-5-6-sol-medium",
          "verified_publish": true,
          "levels_completed": 1,
          "actions": 238,
          "resets": 2,
          "state": "GAME_OVER",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            86.22448979591837,
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            28,
            210,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            26,
            42,
            86,
            108,
            189,
            139,
            424,
            241
          ]
        },
        {
          "session_id": "8f90211c-47a1-4bc5-880a-c5dc465b9020",
          "game_id": "re86",
          "game_version": "8af5384d",
          "published_at": "2026-06-09T21:52:19.461019Z",
          "score": 0,
          "model": "gpt-5-6-sol",
          "runner": "benchmark_agent",
          "config": "openai-gpt-5-6-sol-low",
          "verified_publish": true,
          "levels_completed": 0,
          "actions": 130,
          "resets": 1,
          "state": "GAME_OVER",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            0,
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            130,
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            26,
            42,
            86,
            108,
            189,
            139,
            424,
            241
          ]
        },
        {
          "session_id": "8461c613-b29b-48c9-9342-91f0709c244c",
          "game_id": "re86",
          "game_version": "8af5384d",
          "published_at": "2026-05-29T20:31:11.571615Z",
          "score": 2.7777777777777777,
          "model": "claude-opus-4-8",
          "runner": "benchmark_agent",
          "config": "anthropic-opus-4-8-high",
          "verified_publish": true,
          "levels_completed": 1,
          "actions": 235,
          "resets": 2,
          "state": "GAME_OVER",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            108.16000000000001,
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            25,
            210,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            26,
            42,
            86,
            108,
            189,
            139,
            424,
            241
          ]
        },
        {
          "session_id": "22d0bd59-c0b9-4c05-91d2-2223b86d161c",
          "game_id": "re86",
          "game_version": "8af5384d",
          "published_at": "2026-05-01T05:08:59.409467Z",
          "score": 0.24248163452708912,
          "model": "claude-opus-4-7",
          "runner": "benchmark_agent",
          "config": "anthropic-opus-4-7-high",
          "verified_publish": true,
          "levels_completed": 1,
          "actions": 298,
          "resets": 7,
          "state": "GAME_OVER",
          "completed": false,
          "number_of_levels": 8,
          "level_scores": [
            8.729338842975208,
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            88,
            210,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            26,
            42,
            86,
            108,
            189,
            139,
            424,
            241
          ]
        },
        {
          "session_id": "dc1eff8a-8b48-4d46-aa25-11f1a2730b67",
          "game_id": "re86",
          "game_version": "8af5384d",
          "published_at": "2026-04-30T23:23:49.730755Z",
          "score": 0,
          "model": "gpt-5.5-2026-04-23",
          "runner": "benchmark_agent",
          "config": "openai-gpt-5-5-2026-04-23-high",
          "verified_publish": true,
          "levels_completed": 0,
          "actions": 130,
          "resets": 2,
          "state": "GAME_OVER",
          "completed": false,
          "number_of_levels": 8,
          "level_scores": [
            0,
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            130,
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            26,
            42,
            86,
            108,
            189,
            139,
            424,
            241
          ]
        },
        {
          "session_id": "a88e10a6-b3c8-4de7-be15-2593b2363a07",
          "game_id": "re86",
          "game_version": "8af5384d",
          "published_at": "2026-04-14T17:44:08.882938Z",
          "score": 0,
          "model": "grok-4.20-beta-0309-reasoning",
          "runner": "benchmark_agent",
          "config": "xai-grok-4-20-beta-0309-reasoning",
          "verified_publish": true,
          "levels_completed": 0,
          "actions": 87,
          "resets": 7,
          "state": "NOT_FINISHED",
          "completed": false,
          "number_of_levels": 8,
          "level_scores": [
            0,
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            87,
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            26,
            42,
            86,
            108,
            189,
            139,
            424,
            241
          ]
        },
        {
          "session_id": "80e9023e-bc4f-4556-b589-03b410359e08",
          "game_id": "re86",
          "game_version": "8af5384d",
          "published_at": "2026-04-14T17:44:08.497332Z",
          "score": 0,
          "model": "gpt-5.4-2026-03-05",
          "runner": "benchmark_agent",
          "config": "openai-gpt-5-4-2026-03-05-high",
          "verified_publish": true,
          "levels_completed": 0,
          "actions": 130,
          "resets": 1,
          "state": "GAME_OVER",
          "completed": false,
          "number_of_levels": 8,
          "level_scores": [
            0,
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            130,
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            26,
            42,
            86,
            108,
            189,
            139,
            424,
            241
          ]
        },
        {
          "session_id": "35da5409-996f-4072-a038-efd989c3899d",
          "game_id": "re86",
          "game_version": "8af5384d",
          "published_at": "2026-04-14T17:44:07.76451Z",
          "score": 0,
          "model": "gemini-3.1-pro-preview",
          "runner": "benchmark_agent",
          "config": "google-gemini-3-1-pro-preview",
          "verified_publish": true,
          "levels_completed": 0,
          "actions": 101,
          "resets": 1,
          "state": "GAME_OVER",
          "completed": false,
          "number_of_levels": 8,
          "level_scores": [
            0,
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            101,
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            26,
            42,
            86,
            108,
            189,
            139,
            424,
            241
          ]
        },
        {
          "session_id": "54a97703-fe62-4c60-a41c-1ac06257b135",
          "game_id": "re86",
          "game_version": "8af5384d",
          "published_at": "2026-04-14T17:44:07.382155Z",
          "score": 0,
          "model": "claude-opus-4-6",
          "runner": "benchmark_agent",
          "config": "anthropic-opus-4-6-max-effort",
          "verified_publish": true,
          "levels_completed": 0,
          "actions": 130,
          "resets": 3,
          "state": "NOT_FINISHED",
          "completed": false,
          "number_of_levels": 8,
          "level_scores": [
            0,
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            130,
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            26,
            42,
            86,
            108,
            189,
            139,
            424,
            241
          ]
        }
      ]
    },
    {
      "game": "lf52",
      "source": "https://arcprize.org/tasks/lf52",
      "runs": [
        {
          "session_id": "8b41285d-f4bf-4b46-9278-c8bbaa24611f",
          "game_id": "lf52",
          "game_version": "271a04aa",
          "published_at": "2026-09-25T12:12:47.941921Z",
          "score": 1.8181818181818181,
          "model": "gpt-6-sol",
          "runner": "benchmark_agent",
          "config": "openai-gpt-6-sol-max",
          "verified_publish": true,
          "levels_completed": 1,
          "actions": 430,
          "resets": 0,
          "state": "NOT_FINISHED",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            115,
            0,
            0,
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            25,
            405,
            0,
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            32,
            81,
            60,
            71,
            205,
            148,
            244,
            109,
            164,
            225
          ]
        },
        {
          "session_id": "32d85c47-6217-483f-9a96-c945d74bc146",
          "game_id": "lf52",
          "game_version": "271a04aa",
          "published_at": "2026-09-25T07:24:51.844423Z",
          "score": 0.1250885636803401,
          "model": "gpt-6-sol",
          "runner": "benchmark_agent",
          "config": "openai-gpt-6-sol-xhigh",
          "verified_publish": true,
          "levels_completed": 1,
          "actions": 527,
          "resets": 1,
          "state": "NOT_FINISHED",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            6.879871002418706,
            0,
            0,
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            122,
            405,
            0,
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            32,
            81,
            60,
            71,
            205,
            148,
            244,
            109,
            164,
            225
          ]
        },
        {
          "session_id": "d0592589-a10d-4816-9c07-382686e644c5",
          "game_id": "lf52",
          "game_version": "271a04aa",
          "published_at": "2026-09-25T07:21:27.988369Z",
          "score": 0.28377048953180634,
          "model": "gpt-6-sol",
          "runner": "benchmark_agent",
          "config": "openai-gpt-6-sol-none-provider-adapter",
          "verified_publish": true,
          "levels_completed": 1,
          "actions": 486,
          "resets": 1,
          "state": "NOT_FINISHED",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            15.60737692424935,
            0,
            0,
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            81,
            405,
            0,
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            32,
            81,
            60,
            71,
            205,
            148,
            244,
            109,
            164,
            225
          ]
        },
        {
          "session_id": "72c869f1-b034-48ca-bc11-ca85df2c0ea9",
          "game_id": "lf52",
          "game_version": "271a04aa",
          "published_at": "2026-09-25T06:57:49.642311Z",
          "score": 1.107565842842464,
          "model": "gpt-6-sol",
          "runner": "benchmark_agent",
          "config": "openai-gpt-6-sol-low-provider-adapter",
          "verified_publish": true,
          "levels_completed": 1,
          "actions": 446,
          "resets": 1,
          "state": "NOT_FINISHED",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            60.916121356335516,
            0,
            0,
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            41,
            405,
            0,
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            32,
            81,
            60,
            71,
            205,
            148,
            244,
            109,
            164,
            225
          ]
        },
        {
          "session_id": "e849193c-1c4c-4c88-ac2d-5106a6854215",
          "game_id": "lf52",
          "game_version": "271a04aa",
          "published_at": "2026-09-25T06:43:52.88183Z",
          "score": 1.7096585691627015,
          "model": "gpt-6-sol",
          "runner": "benchmark_agent",
          "config": "openai-gpt-6-sol-medium-provider-adapter",
          "verified_publish": true,
          "levels_completed": 1,
          "actions": 438,
          "resets": 1,
          "state": "NOT_FINISHED",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            94.03122130394858,
            0,
            0,
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            33,
            405,
            0,
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            32,
            81,
            60,
            71,
            205,
            148,
            244,
            109,
            164,
            225
          ]
        },
        {
          "session_id": "cc20b09e-af72-404b-aea2-7b9bf23cf2a5",
          "game_id": "lf52",
          "game_version": "271a04aa",
          "published_at": "2026-09-25T06:34:54.811266Z",
          "score": 0,
          "model": "gpt-6-sol",
          "runner": "benchmark_agent",
          "config": "openai-gpt-6-sol-none",
          "verified_publish": true,
          "levels_completed": 0,
          "actions": 160,
          "resets": 2,
          "state": "GAME_OVER",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            0,
            0,
            0,
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            160,
            0,
            0,
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            32,
            81,
            60,
            71,
            205,
            148,
            244,
            109,
            164,
            225
          ]
        },
        {
          "session_id": "e2dae741-fb9f-49d1-bcaa-e4293612e615",
          "game_id": "lf52",
          "game_version": "271a04aa",
          "published_at": "2026-09-25T06:34:47.386174Z",
          "score": 0.10215737623145031,
          "model": "gpt-6-sol",
          "runner": "benchmark_agent",
          "config": "openai-gpt-6-sol-low",
          "verified_publish": true,
          "levels_completed": 1,
          "actions": 540,
          "resets": 1,
          "state": "NOT_FINISHED",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            5.618655692729767,
            0,
            0,
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            135,
            405,
            0,
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            32,
            81,
            60,
            71,
            205,
            148,
            244,
            109,
            164,
            225
          ]
        },
        {
          "session_id": "e7856f6f-3ff1-429b-883a-d1040463dfc9",
          "game_id": "lf52",
          "game_version": "271a04aa",
          "published_at": "2026-09-25T06:07:16.608428Z",
          "score": 1.0069324942229214,
          "model": "gpt-6-sol",
          "runner": "benchmark_agent",
          "config": "openai-gpt-6-sol-high-provider-adapter",
          "verified_publish": true,
          "levels_completed": 1,
          "actions": 448,
          "resets": 3,
          "state": "NOT_FINISHED",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            55.38128718226068,
            0,
            0,
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            43,
            405,
            0,
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            32,
            81,
            60,
            71,
            205,
            148,
            244,
            109,
            164,
            225
          ]
        },
        {
          "session_id": "f9678e78-bf9c-453a-b98a-db0e8dd6d0c6",
          "game_id": "lf52",
          "game_version": "271a04aa",
          "published_at": "2026-09-25T05:40:36.792725Z",
          "score": 0,
          "model": "gpt-6-sol",
          "runner": "benchmark_agent",
          "config": "openai-gpt-6-sol-medium",
          "verified_publish": true,
          "levels_completed": 0,
          "actions": 160,
          "resets": 1,
          "state": "NOT_FINISHED",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            0,
            0,
            0,
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            160,
            0,
            0,
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            32,
            81,
            60,
            71,
            205,
            148,
            244,
            109,
            164,
            225
          ]
        },
        {
          "session_id": "a0a88ba9-d487-4f1c-9de8-723c576774ec",
          "game_id": "lf52",
          "game_version": "271a04aa",
          "published_at": "2026-09-25T05:26:44.919562Z",
          "score": 1.8181818181818181,
          "model": "gpt-6-sol",
          "runner": "benchmark_agent",
          "config": "openai-gpt-6-sol-xhigh-provider-adapter",
          "verified_publish": true,
          "levels_completed": 1,
          "actions": 413,
          "resets": 0,
          "state": "NOT_FINISHED",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            115,
            0,
            0,
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            8,
            405,
            0,
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            32,
            81,
            60,
            71,
            205,
            148,
            244,
            109,
            164,
            225
          ]
        },
        {
          "session_id": "180c4acf-e673-41d5-94b2-9546b9e818f3",
          "game_id": "lf52",
          "game_version": "271a04aa",
          "published_at": "2026-09-25T04:55:16.709637Z",
          "score": 0,
          "model": "gpt-6-sol",
          "runner": "benchmark_agent",
          "config": "openai-gpt-6-sol-high",
          "verified_publish": true,
          "levels_completed": 0,
          "actions": 160,
          "resets": 1,
          "state": "GAME_OVER",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            0,
            0,
            0,
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            160,
            0,
            0,
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            32,
            81,
            60,
            71,
            205,
            148,
            244,
            109,
            164,
            225
          ]
        },
        {
          "session_id": "56cc76c4-fef8-4167-a141-a5f580f5ffc5",
          "game_id": "lf52",
          "game_version": "271a04aa",
          "published_at": "2026-09-21T10:17:08.965426Z",
          "score": 1.8181818181818181,
          "model": "gemini-3.8-flash",
          "runner": "benchmark_agent",
          "config": "google-gemini-3-8-flash-low-provider-adapter",
          "verified_publish": true,
          "levels_completed": 1,
          "actions": 413,
          "resets": 0,
          "state": "NOT_FINISHED",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            115,
            0,
            0,
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            8,
            405,
            0,
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            32,
            81,
            60,
            71,
            205,
            148,
            244,
            109,
            164,
            225
          ]
        },
        {
          "session_id": "a12204e6-59e0-4b27-81d8-6149d71f6111",
          "game_id": "lf52",
          "game_version": "271a04aa",
          "published_at": "2026-09-20T05:44:15.801564Z",
          "score": 1.8181818181818181,
          "model": "gemini-3.8-flash",
          "runner": "benchmark_agent",
          "config": "google-gemini-3-8-flash-medium-provider-adapter",
          "verified_publish": true,
          "levels_completed": 1,
          "actions": 413,
          "resets": 0,
          "state": "NOT_FINISHED",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            115,
            0,
            0,
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            8,
            405,
            0,
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            32,
            81,
            60,
            71,
            205,
            148,
            244,
            109,
            164,
            225
          ]
        },
        {
          "session_id": "d509c75b-e217-4ed7-9ad1-b667e2c6e591",
          "game_id": "lf52",
          "game_version": "271a04aa",
          "published_at": "2026-09-17T18:57:48.224041Z",
          "score": 1.8181818181818181,
          "model": "gemini-3.8-flash",
          "runner": "benchmark_agent",
          "config": "google-gemini-3-8-flash-high-provider-adapter",
          "verified_publish": true,
          "levels_completed": 1,
          "actions": 413,
          "resets": 1,
          "state": "NOT_FINISHED",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            115,
            0,
            0,
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            8,
            405,
            0,
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            32,
            81,
            60,
            71,
            205,
            148,
            244,
            109,
            164,
            225
          ]
        },
        {
          "session_id": "a83579eb-8a1c-4990-80a8-7deb89406a91",
          "game_id": "lf52",
          "game_version": "271a04aa",
          "published_at": "2026-09-16T20:22:01.575738Z",
          "score": 0,
          "model": "gpt-6-luna",
          "runner": "benchmark_agent",
          "config": "openai-gpt-6-luna-none",
          "verified_publish": true,
          "levels_completed": 0,
          "actions": 160,
          "resets": 2,
          "state": "GAME_OVER",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            0,
            0,
            0,
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            160,
            0,
            0,
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            32,
            81,
            60,
            71,
            205,
            148,
            244,
            109,
            164,
            225
          ]
        },
        {
          "session_id": "29ac927d-4480-45a5-8d15-277a6ed4b390",
          "game_id": "lf52",
          "game_version": "271a04aa",
          "published_at": "2026-09-16T19:37:06.879436Z",
          "score": 0,
          "model": "gpt-6-luna",
          "runner": "benchmark_agent",
          "config": "openai-gpt-6-luna-medium-provider-adapter",
          "verified_publish": true,
          "levels_completed": 0,
          "actions": 160,
          "resets": 0,
          "state": "NOT_FINISHED",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            0,
            0,
            0,
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            160,
            0,
            0,
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            32,
            81,
            60,
            71,
            205,
            148,
            244,
            109,
            164,
            225
          ]
        },
        {
          "session_id": "e05fb64d-d6f3-4c03-b835-ddf290592fa5",
          "game_id": "lf52",
          "game_version": "271a04aa",
          "published_at": "2026-09-16T19:31:11.118301Z",
          "score": 0,
          "model": "gpt-6-luna",
          "runner": "benchmark_agent",
          "config": "openai-gpt-6-luna-medium",
          "verified_publish": true,
          "levels_completed": 0,
          "actions": 160,
          "resets": 2,
          "state": "GAME_OVER",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            0,
            0,
            0,
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            160,
            0,
            0,
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            32,
            81,
            60,
            71,
            205,
            148,
            244,
            109,
            164,
            225
          ]
        },
        {
          "session_id": "452790c2-1535-4657-a68e-f6cb0430de3c",
          "game_id": "lf52",
          "game_version": "271a04aa",
          "published_at": "2026-09-16T19:28:40.115372Z",
          "score": 0,
          "model": "gpt-6-luna",
          "runner": "benchmark_agent",
          "config": "openai-gpt-6-luna-low-provider-adapter",
          "verified_publish": true,
          "levels_completed": 0,
          "actions": 160,
          "resets": 1,
          "state": "GAME_OVER",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            0,
            0,
            0,
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            160,
            0,
            0,
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            32,
            81,
            60,
            71,
            205,
            148,
            244,
            109,
            164,
            225
          ]
        },
        {
          "session_id": "90ad87db-f298-44b8-a4d1-ece8600c7cf8",
          "game_id": "lf52",
          "game_version": "271a04aa",
          "published_at": "2026-09-16T19:27:08.757741Z",
          "score": 0,
          "model": "gpt-6-luna",
          "runner": "benchmark_agent",
          "config": "openai-gpt-6-luna-none-provider-adapter",
          "verified_publish": true,
          "levels_completed": 0,
          "actions": 160,
          "resets": 1,
          "state": "NOT_FINISHED",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            0,
            0,
            0,
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            160,
            0,
            0,
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            32,
            81,
            60,
            71,
            205,
            148,
            244,
            109,
            164,
            225
          ]
        },
        {
          "session_id": "2a93cee3-88c7-4b5c-a2b2-b9b1fc001195",
          "game_id": "lf52",
          "game_version": "271a04aa",
          "published_at": "2026-09-16T19:26:14.945637Z",
          "score": 0,
          "model": "gpt-6-luna",
          "runner": "benchmark_agent",
          "config": "openai-gpt-6-luna-low",
          "verified_publish": true,
          "levels_completed": 0,
          "actions": 160,
          "resets": 1,
          "state": "GAME_OVER",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            0,
            0,
            0,
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            160,
            0,
            0,
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            32,
            81,
            60,
            71,
            205,
            148,
            244,
            109,
            164,
            225
          ]
        },
        {
          "session_id": "7c105a53-db6e-4a7c-ae0b-0c2434ad1560",
          "game_id": "lf52",
          "game_version": "271a04aa",
          "published_at": "2026-09-16T18:17:13.738501Z",
          "score": 0.6885422270037654,
          "model": "gpt-6-luna",
          "runner": "benchmark_agent",
          "config": "openai-gpt-6-luna-xhigh-provider-adapter",
          "verified_publish": true,
          "levels_completed": 1,
          "actions": 447,
          "resets": 1,
          "state": "NOT_FINISHED",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            37.8698224852071,
            0,
            0,
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            52,
            395,
            0,
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            32,
            81,
            60,
            71,
            205,
            148,
            244,
            109,
            164,
            225
          ]
        },
        {
          "session_id": "5aaca111-b637-448c-bf7a-d74cb6178c8c",
          "game_id": "lf52",
          "game_version": "271a04aa",
          "published_at": "2026-09-16T17:28:14.734765Z",
          "score": 100,
          "model": "gpt-6-1-sol",
          "runner": "benchmark_agent",
          "config": "openai-gpt-6-1-sol-high-provider-adapter",
          "verified_publish": true,
          "levels_completed": 10,
          "actions": 716,
          "resets": 0,
          "state": "WIN",
          "completed": true,
          "number_of_levels": 0,
          "level_scores": [
            115,
            115,
            115,
            115,
            115,
            115,
            115,
            115,
            115,
            115
          ],
          "level_actions": [
            8,
            36,
            46,
            52,
            106,
            93,
            144,
            73,
            106,
            52
          ],
          "level_baseline_actions": [
            32,
            81,
            60,
            71,
            205,
            148,
            244,
            109,
            164,
            225
          ]
        },
        {
          "session_id": "3cf414d1-c133-4809-b0f3-ce5a97231a54",
          "game_id": "lf52",
          "game_version": "271a04aa",
          "published_at": "2026-09-16T16:00:21.120395Z",
          "score": 0,
          "model": "gpt-6-luna",
          "runner": "benchmark_agent",
          "config": "openai-gpt-6-luna-high",
          "verified_publish": true,
          "levels_completed": 0,
          "actions": 160,
          "resets": 1,
          "state": "GAME_OVER",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            0,
            0,
            0,
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            160,
            0,
            0,
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            32,
            81,
            60,
            71,
            205,
            148,
            244,
            109,
            164,
            225
          ]
        },
        {
          "session_id": "ff627a47-4bb9-4590-9204-ee5389698cf7",
          "game_id": "lf52",
          "game_version": "271a04aa",
          "published_at": "2026-09-16T09:54:58.382954Z",
          "score": 0,
          "model": "gpt-6-luna",
          "runner": "benchmark_agent",
          "config": "openai-gpt-6-luna-high-provider-adapter",
          "verified_publish": true,
          "levels_completed": 0,
          "actions": 155,
          "resets": 2,
          "state": "NOT_FINISHED",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            0,
            0,
            0,
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            155,
            0,
            0,
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            32,
            81,
            60,
            71,
            205,
            148,
            244,
            109,
            164,
            225
          ]
        },
        {
          "session_id": "83f51a13-da32-4035-b0ff-3da9e8d274a4",
          "game_id": "lf52",
          "game_version": "271a04aa",
          "published_at": "2026-09-16T08:29:09.136977Z",
          "score": 0,
          "model": "gpt-6-luna",
          "runner": "benchmark_agent",
          "config": "openai-gpt-6-luna-max",
          "verified_publish": true,
          "levels_completed": 0,
          "actions": 160,
          "resets": 1,
          "state": "GAME_OVER",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            0,
            0,
            0,
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            160,
            0,
            0,
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            32,
            81,
            60,
            71,
            205,
            148,
            244,
            109,
            164,
            225
          ]
        },
        {
          "session_id": "ef54371b-5e0b-428b-b105-944ca0853fbd",
          "game_id": "lf52",
          "game_version": "271a04aa",
          "published_at": "2026-09-16T08:08:26.945158Z",
          "score": 0,
          "model": "gpt-6-luna",
          "runner": "benchmark_agent",
          "config": "openai-gpt-6-luna-xhigh",
          "verified_publish": true,
          "levels_completed": 0,
          "actions": 160,
          "resets": 1,
          "state": "GAME_OVER",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            0,
            0,
            0,
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            160,
            0,
            0,
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            32,
            81,
            60,
            71,
            205,
            148,
            244,
            109,
            164,
            225
          ]
        },
        {
          "session_id": "77ed9d0f-7424-414d-a755-971dcf8c4ded",
          "game_id": "lf52",
          "game_version": "271a04aa",
          "published_at": "2026-09-16T07:30:29.167432Z",
          "score": 1.8181818181818181,
          "model": "gpt-6-1-sol",
          "runner": "benchmark_agent",
          "config": "openai-gpt-6-1-sol-low",
          "verified_publish": true,
          "levels_completed": 1,
          "actions": 419,
          "resets": 1,
          "state": "NOT_FINISHED",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            115,
            0,
            0,
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            14,
            405,
            0,
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            32,
            81,
            60,
            71,
            205,
            148,
            244,
            109,
            164,
            225
          ]
        },
        {
          "session_id": "cb21f97a-2ab9-49f5-82ee-cb8d5266546e",
          "game_id": "lf52",
          "game_version": "271a04aa",
          "published_at": "2026-09-16T07:15:25.983671Z",
          "score": 5.454545454545454,
          "model": "gpt-6-1-sol",
          "runner": "benchmark_agent",
          "config": "openai-gpt-6-1-sol-max",
          "verified_publish": true,
          "levels_completed": 2,
          "actions": 355,
          "resets": 1,
          "state": "NOT_FINISHED",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            115,
            115,
            0,
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            8,
            47,
            300,
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            32,
            81,
            60,
            71,
            205,
            148,
            244,
            109,
            164,
            225
          ]
        },
        {
          "session_id": "f9f9d28b-7039-408e-b104-047826046690",
          "game_id": "lf52",
          "game_version": "271a04aa",
          "published_at": "2026-09-16T06:21:21.506163Z",
          "score": 3.863960652902389,
          "model": "gpt-6-1-sol",
          "runner": "benchmark_agent",
          "config": "openai-gpt-6-1-sol-medium",
          "verified_publish": true,
          "levels_completed": 2,
          "actions": 429,
          "resets": 0,
          "state": "NOT_FINISHED",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            115,
            48.75891795481569,
            0,
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            13,
            116,
            300,
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            32,
            81,
            60,
            71,
            205,
            148,
            244,
            109,
            164,
            225
          ]
        },
        {
          "session_id": "a4ed293f-baf4-4b4f-9a7a-0a559fd8c7f4",
          "game_id": "lf52",
          "game_version": "271a04aa",
          "published_at": "2026-09-16T05:54:31.567887Z",
          "score": 2.265183611129557,
          "model": "gpt-6-1-sol",
          "runner": "benchmark_agent",
          "config": "openai-gpt-6-1-sol-xhigh",
          "verified_publish": true,
          "levels_completed": 2,
          "actions": 678,
          "resets": 3,
          "state": "NOT_FINISHED",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            115,
            4.792549306062821,
            0,
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            8,
            370,
            300,
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            32,
            81,
            60,
            71,
            205,
            148,
            244,
            109,
            164,
            225
          ]
        },
        {
          "session_id": "c972f6e7-5c08-47f9-ac32-acbee9473861",
          "game_id": "lf52",
          "game_version": "271a04aa",
          "published_at": "2026-09-16T05:47:29.284781Z",
          "score": 100,
          "model": "gpt-6-1-sol",
          "runner": "benchmark_agent",
          "config": "openai-gpt-6-1-sol-xhigh-provider-adapter",
          "verified_publish": true,
          "levels_completed": 10,
          "actions": 761,
          "resets": 0,
          "state": "WIN",
          "completed": true,
          "number_of_levels": 0,
          "level_scores": [
            115,
            115,
            115,
            115,
            115,
            115,
            115,
            115,
            115,
            115
          ],
          "level_actions": [
            8,
            41,
            47,
            54,
            93,
            116,
            146,
            75,
            104,
            77
          ],
          "level_baseline_actions": [
            32,
            81,
            60,
            71,
            205,
            148,
            244,
            109,
            164,
            225
          ]
        },
        {
          "session_id": "d3a02d55-f98e-48f6-82c0-4dca604881ef",
          "game_id": "lf52",
          "game_version": "271a04aa",
          "published_at": "2026-09-16T05:16:16.422421Z",
          "score": 100,
          "model": "gpt-6-1-sol",
          "runner": "benchmark_agent",
          "config": "openai-gpt-6-1-sol-max-provider-adapter",
          "verified_publish": true,
          "levels_completed": 10,
          "actions": 835,
          "resets": 0,
          "state": "WIN",
          "completed": true,
          "number_of_levels": 0,
          "level_scores": [
            115,
            115,
            115,
            115,
            115,
            75.79238754325259,
            115,
            115,
            115,
            115
          ],
          "level_actions": [
            8,
            42,
            48,
            52,
            123,
            170,
            144,
            68,
            120,
            60
          ],
          "level_baseline_actions": [
            32,
            81,
            60,
            71,
            205,
            148,
            244,
            109,
            164,
            225
          ]
        },
        {
          "session_id": "42b1e78a-3dbc-4d93-8cf8-ef74560bde28",
          "game_id": "lf52",
          "game_version": "271a04aa",
          "published_at": "2026-09-16T04:29:32.849008Z",
          "score": 1.8181818181818181,
          "model": "gpt-6-1-sol",
          "runner": "benchmark_agent",
          "config": "openai-gpt-6-1-sol-high",
          "verified_publish": true,
          "levels_completed": 1,
          "actions": 413,
          "resets": 1,
          "state": "GAME_OVER",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            115,
            0,
            0,
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            8,
            405,
            0,
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            32,
            81,
            60,
            71,
            205,
            148,
            244,
            109,
            164,
            225
          ]
        },
        {
          "session_id": "fc9a0c12-dde7-4d5e-9b25-6d9db7e619d6",
          "game_id": "lf52",
          "game_version": "271a04aa",
          "published_at": "2026-09-16T03:26:57.588792Z",
          "score": 0,
          "model": "gpt-6-luna",
          "runner": "benchmark_agent",
          "config": "openai-gpt-6-luna-max-provider-adapter",
          "verified_publish": true,
          "levels_completed": 0,
          "actions": 160,
          "resets": 2,
          "state": "GAME_OVER",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            0,
            0,
            0,
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            160,
            0,
            0,
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            32,
            81,
            60,
            71,
            205,
            148,
            244,
            109,
            164,
            225
          ]
        },
        {
          "session_id": "a5e7119c-2938-421a-a266-c53fef41e3b7",
          "game_id": "lf52",
          "game_version": "271a04aa",
          "published_at": "2026-09-12T06:46:46.273611Z",
          "score": 1.8181818181818181,
          "model": "gemini-3.8-flash",
          "runner": "benchmark_agent",
          "config": "google-gemini-3-8-flash-low",
          "verified_publish": true,
          "levels_completed": 1,
          "actions": 421,
          "resets": 1,
          "state": "NOT_FINISHED",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            115,
            0,
            0,
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            16,
            405,
            0,
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            32,
            81,
            60,
            71,
            205,
            148,
            244,
            109,
            164,
            225
          ]
        },
        {
          "session_id": "973f35c7-f6a7-48bf-98b1-f25bb74d1fe0",
          "game_id": "lf52",
          "game_version": "271a04aa",
          "published_at": "2026-09-11T00:53:57.656451Z",
          "score": 1.8181818181818181,
          "model": "gemini-3.8-flash",
          "runner": "benchmark_agent",
          "config": "google-gemini-3-8-flash-high",
          "verified_publish": true,
          "levels_completed": 1,
          "actions": 413,
          "resets": 2,
          "state": "NOT_FINISHED",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            115,
            0,
            0,
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            8,
            405,
            0,
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            32,
            81,
            60,
            71,
            205,
            148,
            244,
            109,
            164,
            225
          ]
        },
        {
          "session_id": "8c22d6a1-023d-4841-9330-a17ad051cc93",
          "game_id": "lf52",
          "game_version": "271a04aa",
          "published_at": "2026-09-10T23:34:52.891504Z",
          "score": 1.8181818181818181,
          "model": "gemini-3.8-flash",
          "runner": "benchmark_agent",
          "config": "google-gemini-3-8-flash-medium",
          "verified_publish": true,
          "levels_completed": 1,
          "actions": 417,
          "resets": 1,
          "state": "NOT_FINISHED",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            115,
            0,
            0,
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            12,
            405,
            0,
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            32,
            81,
            60,
            71,
            205,
            148,
            244,
            109,
            164,
            225
          ]
        },
        {
          "session_id": "7f8fa582-3eda-48c7-b77b-6b30fb7d55b6",
          "game_id": "lf52",
          "game_version": "271a04aa",
          "published_at": "2026-08-11T13:51:52.707903Z",
          "score": 0,
          "model": "grok-4.6",
          "runner": "benchmark_agent",
          "config": "xai-grok-4-6-xhigh",
          "verified_publish": true,
          "levels_completed": 0,
          "actions": 160,
          "resets": 6,
          "state": "NOT_FINISHED",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            0,
            0,
            0,
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            160,
            0,
            0,
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            32,
            81,
            60,
            71,
            205,
            148,
            244,
            109,
            164,
            225
          ]
        },
        {
          "session_id": "5d0b7a9c-5e27-4100-a97c-53cf95f7a3bf",
          "game_id": "lf52",
          "game_version": "271a04aa",
          "published_at": "2026-07-21T14:11:47.594857Z",
          "score": 1.8181818181818181,
          "model": "claude-opus-5",
          "runner": "benchmark_agent",
          "config": "anthropic-claude-opus-5-high",
          "verified_publish": true,
          "levels_completed": 1,
          "actions": 419,
          "resets": 8,
          "state": "NOT_FINISHED",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            115,
            0,
            0,
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            14,
            405,
            0,
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            32,
            81,
            60,
            71,
            205,
            148,
            244,
            109,
            164,
            225
          ]
        },
        {
          "session_id": "aa2930da-d794-4677-8038-a240b779c5b0",
          "game_id": "lf52",
          "game_version": "271a04aa",
          "published_at": "2026-06-24T11:00:31.280889Z",
          "score": 0,
          "model": "gpt-5-6-luna",
          "runner": "benchmark_agent",
          "config": "openai-gpt-5-6-luna-low",
          "verified_publish": true,
          "levels_completed": 0,
          "actions": 160,
          "resets": 2,
          "state": "GAME_OVER",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            0,
            0,
            0,
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            160,
            0,
            0,
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            32,
            81,
            60,
            71,
            205,
            148,
            244,
            109,
            164,
            225
          ]
        },
        {
          "session_id": "312175c4-d552-486c-85a3-5d7df08ea684",
          "game_id": "lf52",
          "game_version": "271a04aa",
          "published_at": "2026-06-24T10:40:26.694908Z",
          "score": 0,
          "model": "gpt-5-6-luna",
          "runner": "benchmark_agent",
          "config": "openai-gpt-5-6-luna-medium",
          "verified_publish": true,
          "levels_completed": 0,
          "actions": 160,
          "resets": 1,
          "state": "GAME_OVER",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            0,
            0,
            0,
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            160,
            0,
            0,
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            32,
            81,
            60,
            71,
            205,
            148,
            244,
            109,
            164,
            225
          ]
        },
        {
          "session_id": "520a065b-cfd8-4aa1-af0d-c26daacd12ea",
          "game_id": "lf52",
          "game_version": "271a04aa",
          "published_at": "2026-06-24T04:34:02.433496Z",
          "score": 0,
          "model": "gpt-5-6-luna",
          "runner": "benchmark_agent",
          "config": "openai-gpt-5-6-luna-high",
          "verified_publish": true,
          "levels_completed": 0,
          "actions": 160,
          "resets": 5,
          "state": "NOT_FINISHED",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            0,
            0,
            0,
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            160,
            0,
            0,
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            32,
            81,
            60,
            71,
            205,
            148,
            244,
            109,
            164,
            225
          ]
        },
        {
          "session_id": "16426f45-255d-42c7-9e9a-3addecb8dc98",
          "game_id": "lf52",
          "game_version": "271a04aa",
          "published_at": "2026-06-24T03:23:06.124252Z",
          "score": 0,
          "model": "gpt-5-6-luna",
          "runner": "benchmark_agent",
          "config": "openai-gpt-5-6-luna-xhigh",
          "verified_publish": true,
          "levels_completed": 0,
          "actions": 160,
          "resets": 16,
          "state": "NOT_FINISHED",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            0,
            0,
            0,
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            160,
            0,
            0,
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            32,
            81,
            60,
            71,
            205,
            148,
            244,
            109,
            164,
            225
          ]
        },
        {
          "session_id": "82e1779e-3aeb-4bd6-a29c-5172f97cc66e",
          "game_id": "lf52",
          "game_version": "271a04aa",
          "published_at": "2026-06-23T04:34:34.923396Z",
          "score": 0,
          "model": "gpt-5-6-luna",
          "runner": "benchmark_agent",
          "config": "openai-gpt-5-6-luna-max",
          "verified_publish": true,
          "levels_completed": 0,
          "actions": 160,
          "resets": 23,
          "state": "NOT_FINISHED",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            0,
            0,
            0,
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            160,
            0,
            0,
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            32,
            81,
            60,
            71,
            205,
            148,
            244,
            109,
            164,
            225
          ]
        },
        {
          "session_id": "99d53608-5614-4962-b48f-10298b8bf9cb",
          "game_id": "lf52",
          "game_version": "271a04aa",
          "published_at": "2026-06-20T12:22:32.477797Z",
          "score": 0.15670551147362863,
          "model": "gpt-5-6-terra",
          "runner": "benchmark_agent",
          "config": "openai-gpt-5-6-terra-low",
          "verified_publish": true,
          "levels_completed": 1,
          "actions": 514,
          "resets": 0,
          "state": "NOT_FINISHED",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            8.618803131049575,
            0,
            0,
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            109,
            405,
            0,
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            32,
            81,
            60,
            71,
            205,
            148,
            244,
            109,
            164,
            225
          ]
        },
        {
          "session_id": "d72eb986-5e5d-436b-bb5d-dd05e123225c",
          "game_id": "lf52",
          "game_version": "271a04aa",
          "published_at": "2026-06-20T09:35:28.186144Z",
          "score": 0,
          "model": "gpt-5-6-terra",
          "runner": "benchmark_agent",
          "config": "openai-gpt-5-6-terra-medium",
          "verified_publish": true,
          "levels_completed": 0,
          "actions": 160,
          "resets": 0,
          "state": "NOT_FINISHED",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            0,
            0,
            0,
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            160,
            0,
            0,
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            32,
            81,
            60,
            71,
            205,
            148,
            244,
            109,
            164,
            225
          ]
        },
        {
          "session_id": "2485f622-c70b-469c-b40f-0c72a9c00b49",
          "game_id": "lf52",
          "game_version": "271a04aa",
          "published_at": "2026-06-20T08:52:04.634289Z",
          "score": 0,
          "model": "gpt-5-6-terra",
          "runner": "benchmark_agent",
          "config": "openai-gpt-5-6-terra-high",
          "verified_publish": true,
          "levels_completed": 0,
          "actions": 160,
          "resets": 1,
          "state": "GAME_OVER",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            0,
            0,
            0,
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            160,
            0,
            0,
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            32,
            81,
            60,
            71,
            205,
            148,
            244,
            109,
            164,
            225
          ]
        },
        {
          "session_id": "f55d7475-d3cc-41b7-839b-586a9fb195b7",
          "game_id": "lf52",
          "game_version": "271a04aa",
          "published_at": "2026-06-20T04:20:21.690721Z",
          "score": 0,
          "model": "gpt-5-6-terra",
          "runner": "benchmark_agent",
          "config": "openai-gpt-5-6-terra-xhigh",
          "verified_publish": true,
          "levels_completed": 0,
          "actions": 160,
          "resets": 4,
          "state": "GAME_OVER",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            0,
            0,
            0,
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            160,
            0,
            0,
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            32,
            81,
            60,
            71,
            205,
            148,
            244,
            109,
            164,
            225
          ]
        },
        {
          "session_id": "ca970049-cb2f-4e94-b325-a59a25ad1c64",
          "game_id": "lf52",
          "game_version": "271a04aa",
          "published_at": "2026-06-19T00:25:01.277341Z",
          "score": 0,
          "model": "gpt-5-6-terra",
          "runner": "benchmark_agent",
          "config": "openai-gpt-5-6-terra-max",
          "verified_publish": true,
          "levels_completed": 0,
          "actions": 160,
          "resets": 4,
          "state": "GAME_OVER",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            0,
            0,
            0,
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            160,
            0,
            0,
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            32,
            81,
            60,
            71,
            205,
            148,
            244,
            109,
            164,
            225
          ]
        },
        {
          "session_id": "f79b95c8-6350-4334-a444-33936a428324",
          "game_id": "lf52",
          "game_version": "271a04aa",
          "published_at": "2026-06-13T12:51:22.101835Z",
          "score": 1.8181818181818181,
          "model": "gpt-5-6-sol",
          "runner": "benchmark_agent",
          "config": "openai-gpt-5-6-sol-max",
          "verified_publish": true,
          "levels_completed": 1,
          "actions": 413,
          "resets": 2,
          "state": "NOT_FINISHED",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            115,
            0,
            0,
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            8,
            405,
            0,
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            32,
            81,
            60,
            71,
            205,
            148,
            244,
            109,
            164,
            225
          ]
        },
        {
          "session_id": "6d1d0852-6862-489c-8b23-5e39d6b37280",
          "game_id": "lf52",
          "game_version": "271a04aa",
          "published_at": "2026-06-12T00:32:20.100542Z",
          "score": 1.8181818181818181,
          "model": "gpt-5-6-sol",
          "runner": "benchmark_agent",
          "config": "openai-gpt-5-6-sol-xhigh",
          "verified_publish": true,
          "levels_completed": 1,
          "actions": 130,
          "resets": 2,
          "state": "NOT_FINISHED",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            115,
            0,
            0,
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            10,
            120,
            0,
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            32,
            81,
            60,
            71,
            205,
            148,
            244,
            109,
            164,
            225
          ]
        },
        {
          "session_id": "09804a54-0ac0-4940-a6f5-961e55623a02",
          "game_id": "lf52",
          "game_version": "271a04aa",
          "published_at": "2026-06-10T14:11:41.676226Z",
          "score": 0,
          "model": "gpt-5-6-sol",
          "runner": "benchmark_agent",
          "config": "openai-gpt-5-6-sol-high",
          "verified_publish": true,
          "levels_completed": 0,
          "actions": 160,
          "resets": 2,
          "state": "GAME_OVER",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            0,
            0,
            0,
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            160,
            0,
            0,
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            32,
            81,
            60,
            71,
            205,
            148,
            244,
            109,
            164,
            225
          ]
        },
        {
          "session_id": "238d6b73-7338-4b81-b0ee-392e6f1aa5fa",
          "game_id": "lf52",
          "game_version": "271a04aa",
          "published_at": "2026-06-10T09:59:38.68718Z",
          "score": 0.09104690605008468,
          "model": "gpt-5-6-sol",
          "runner": "benchmark_agent",
          "config": "openai-gpt-5-6-sol-medium",
          "verified_publish": true,
          "levels_completed": 1,
          "actions": 548,
          "resets": 2,
          "state": "GAME_OVER",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            5.007579832754658,
            0,
            0,
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            143,
            405,
            0,
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            32,
            81,
            60,
            71,
            205,
            148,
            244,
            109,
            164,
            225
          ]
        },
        {
          "session_id": "d15ae60e-a170-4953-872a-f917cb06adf7",
          "game_id": "lf52",
          "game_version": "271a04aa",
          "published_at": "2026-06-09T22:51:41.428132Z",
          "score": 0,
          "model": "gpt-5-6-sol",
          "runner": "benchmark_agent",
          "config": "openai-gpt-5-6-sol-low",
          "verified_publish": true,
          "levels_completed": 0,
          "actions": 160,
          "resets": 2,
          "state": "GAME_OVER",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            0,
            0,
            0,
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            160,
            0,
            0,
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            32,
            81,
            60,
            71,
            205,
            148,
            244,
            109,
            164,
            225
          ]
        },
        {
          "session_id": "8a5dc654-9240-40ce-b0cc-03308f6ab6dc",
          "game_id": "lf52",
          "game_version": "271a04aa",
          "published_at": "2026-05-29T17:29:45.539007Z",
          "score": 0,
          "model": "claude-opus-4-8",
          "runner": "benchmark_agent",
          "config": "anthropic-opus-4-8-high",
          "verified_publish": true,
          "levels_completed": 0,
          "actions": 160,
          "resets": 3,
          "state": "GAME_OVER",
          "completed": false,
          "number_of_levels": 0,
          "level_scores": [
            0,
            0,
            0,
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            160,
            0,
            0,
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            32,
            81,
            60,
            71,
            205,
            148,
            244,
            109,
            164,
            225
          ]
        },
        {
          "session_id": "b2e1c25e-9637-423f-8dfe-0a89bd84fa7f",
          "game_id": "lf52",
          "game_version": "271a04aa",
          "published_at": "2026-05-01T16:12:09.638434Z",
          "score": 0,
          "model": "claude-opus-4-7",
          "runner": "benchmark_agent",
          "config": "anthropic-opus-4-7-high",
          "verified_publish": true,
          "levels_completed": 0,
          "actions": 160,
          "resets": 2,
          "state": "NOT_FINISHED",
          "completed": false,
          "number_of_levels": 10,
          "level_scores": [
            0,
            0,
            0,
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            160,
            0,
            0,
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            32,
            81,
            60,
            71,
            205,
            148,
            244,
            109,
            164,
            225
          ]
        },
        {
          "session_id": "dfa07b1b-80d2-448c-a827-b45f1e635473",
          "game_id": "lf52",
          "game_version": "271a04aa",
          "published_at": "2026-04-30T23:25:25.003372Z",
          "score": 0.5171717171717172,
          "model": "gpt-5.5-2026-04-23",
          "runner": "benchmark_agent",
          "config": "openai-gpt-5-5-2026-04-23-high",
          "verified_publish": true,
          "levels_completed": 1,
          "actions": 465,
          "resets": 2,
          "state": "GAME_OVER",
          "completed": false,
          "number_of_levels": 10,
          "level_scores": [
            28.444444444444443,
            0,
            0,
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            60,
            405,
            0,
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            32,
            81,
            60,
            71,
            205,
            148,
            244,
            109,
            164,
            225
          ]
        },
        {
          "session_id": "5454a987-36f7-4782-a11a-524139198f2c",
          "game_id": "lf52",
          "game_version": "271a04aa",
          "published_at": "2026-04-14T17:44:14.342141Z",
          "score": 0,
          "model": "grok-4.20-beta-0309-reasoning",
          "runner": "benchmark_agent",
          "config": "xai-grok-4-20-beta-0309-reasoning",
          "verified_publish": true,
          "levels_completed": 0,
          "actions": 136,
          "resets": 2,
          "state": "GAME_OVER",
          "completed": false,
          "number_of_levels": 10,
          "level_scores": [
            0,
            0,
            0,
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            136,
            0,
            0,
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            32,
            81,
            60,
            71,
            205,
            148,
            244,
            109,
            164,
            225
          ]
        },
        {
          "session_id": "05a103cf-ca63-45bd-be64-4a94979c5af2",
          "game_id": "lf52",
          "game_version": "271a04aa",
          "published_at": "2026-04-14T17:44:13.802781Z",
          "score": 0,
          "model": "gpt-5.4-2026-03-05",
          "runner": "benchmark_agent",
          "config": "openai-gpt-5-4-2026-03-05-high",
          "verified_publish": true,
          "levels_completed": 0,
          "actions": 160,
          "resets": 0,
          "state": "NOT_FINISHED",
          "completed": false,
          "number_of_levels": 10,
          "level_scores": [
            0,
            0,
            0,
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            160,
            0,
            0,
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            32,
            81,
            60,
            71,
            205,
            148,
            244,
            109,
            164,
            225
          ]
        },
        {
          "session_id": "0f982675-1f2d-4801-a715-5fbf24b0a11a",
          "game_id": "lf52",
          "game_version": "271a04aa",
          "published_at": "2026-04-14T17:44:13.478523Z",
          "score": 1.8181818181818181,
          "model": "gemini-3.1-pro-preview",
          "runner": "benchmark_agent",
          "config": "google-gemini-3-1-pro-preview",
          "verified_publish": true,
          "levels_completed": 1,
          "actions": 136,
          "resets": 0,
          "state": "NOT_FINISHED",
          "completed": false,
          "number_of_levels": 10,
          "level_scores": [
            115,
            0,
            0,
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            15,
            121,
            0,
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            32,
            81,
            60,
            71,
            205,
            148,
            244,
            109,
            164,
            225
          ]
        },
        {
          "session_id": "2eaf7ff0-1cf9-45a4-b7d3-f6d7d5cba0f3",
          "game_id": "lf52",
          "game_version": "271a04aa",
          "published_at": "2026-04-14T17:44:11.183087Z",
          "score": 0,
          "model": "claude-opus-4-6",
          "runner": "benchmark_agent",
          "config": "anthropic-opus-4-6-max-effort",
          "verified_publish": true,
          "levels_completed": 0,
          "actions": 160,
          "resets": 2,
          "state": "GAME_OVER",
          "completed": false,
          "number_of_levels": 10,
          "level_scores": [
            0,
            0,
            0,
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_actions": [
            160,
            0,
            0,
            0,
            0,
            0,
            0,
            0,
            0,
            0
          ],
          "level_baseline_actions": [
            32,
            81,
            60,
            71,
            205,
            148,
            244,
            109,
            164,
            225
          ]
        }
      ]
    }
  ]
}
