Reference page for the ARC Prize 2026 Paper Track write-up "Abundant Tokens and Generalized Tools". Team Turducken, 9 November 2026.

Evidence notes

These notes list the figures used in the write-up and where each comes from. They are our own records, not independently reproduced results. Public leaderboard scores are public scores only; we have not seen the private leaderboard.

1. Notebooks and scores

• Earlier system: public score 56.02 (submitted 23 October), notebook version 361507979.
• Later harness: public score 53.70 (submitted 28 October), notebook version 363005415. It is the earlier system plus expert-kernel rounding and other kernel and serving changes, scorecard keep-alive and a fair-floor scheduling queue, with no cache splice. The same package resubmitted unchanged scored 50.34 (30 October).

2. Configuration

Qwen3.8-Flash-Next with Intel four-bit AutoRound weights, Daniel Franzen's SGLang serving fork and Milestone 2 harness; one RTX PRO 6000 (96 GB); 110 hidden games per scored run in about nine hours (the recorded 56.02 run took 9 h 13 min). Sampling: temperature 0.5, top-p 0.95, top-k 20. Context window 140,000 tokens; a trim keeps the system prompt, the notes and the last 52,000 tokens. Tool output is capped near 3,072 tokens and each Python call at 30 seconds.

3. Throughput and token budget

Whole-run throughput on the local 50-game board rose from 605 tokens per second (Franzen's notebook) to 2,150 generated tokens per second (the 56.02 configuration). The budget in the write-up is an extrapolation: 2,150 × 9 × 3,600 ≈ 69.7 million generated tokens, or about 633,000 per game if shared equally across 110 games. It is not a measured Kaggle allocation, because games share the GPU and use different amounts of time.

4. Within-run score climbs (Figure 1)

All finished runs on our 50-game board (25 public games and 25 copycat variants) that reached 200 minutes: 321 runs. Selection is by completion only, not by score; failed or shorter runs are excluded. Many configurations and repeats are pooled, so this is not a treatment comparison. Scores are linearly interpolated to whole minutes, and unplayed games count as zero. Median board score: 45.18 at minute 100 and 60.93 at minute 200. Median actions per run: 12,415 and 22,105. Action counts are not token counts, and the curves do not separate game difficulty from scheduling.

5. Memory-bundle comparison (Figure 2)

Same 50-game board, 14 October, 200-minute runs: four base runs against six runs with notes or a dictionary store, retained functions and a batching instruction switched on together. Mean score at minute 200: 55.85 against 59.04, a gain of 3.2 points (95% interval −2.0 to +8.3; permutation p ≈ 0.09). The gain was larger on the copycat games (+4.9) than on the official games (+1.5), a split made after the fact. Because the three additions were enabled together, this is not an ablation of any one of them, and it is not a Kaggle score curve.

6. Fourteen paired local and Kaggle comparisons (Figure 3)

Each row compares a package against its own predecessor or control, in score points. Local game sets differ between rows (including a 32-game set and a copycat subset), and packages bundle several changes. Twelve rows improved locally; three of those also improved on Kaggle; a fourth Kaggle gain (the speed package) had a negative local change. Our local repeat standard deviation is 1.9 points over eight identical controls; the unchanged package moved from 53.70 to 50.34 on resubmission. The earlier engineering draft has been retired.

Three further Kaggle results had no matching local test and are excluded from the fourteen: two reinforcement-learning adapters (−1.7 and −8.1) and the unchanged resubmission (−3.4).