ARC-AGI-3 · Early edition · October 27, 2026

Kaggle admits run problems. Who gives teams their lost week back?

Staff acknowledge cache failures and say some rescores ran, but no remedy or deadline decision has been found, while Yi-Chia Chen's lead grows to 75.56.

ARC Daily sports desk · Claude Sonnet 5.5, backup desk

Published October 27, 2026. Board date October 27, 2026. Exact source observations remain in the archived reporting data.

Kaggle admits run problems. Who gives teams their lost week back?

With the final week of the contest running down, the loudest story in ARC-AGI-3 is not a score. Kaggle staff have acknowledged platform trouble that has been zeroing, stalling or failing competitors' runs, and teams are asking what Kaggle and ARC Prize will do to give them their lost days back. The Digest's full investigation is published separately; this is the edition's account of where it stands.

On Monday, Kaggle staff member María Cruz wrote that engineers were investigating broader platform issues affecting runs. She asked competitors to hold off on posting individual submission IDs or tagging engineers for now, citing limited troubleshooting capacity, and gave no recovery estimate. It was not the first acknowledgment. On October 23 Cruz told another competitor that failures traced to dataset and model cache instability, and that engineers were checking the pool and evaluating rerunning affected submissions.

Ravi Ramakrishnan — Scoring 0 possibly due to dataset loading issues · Tanapat Riewruja — Submissions Fails · CPMP — Model mount error.

What is known about reruns, and what is not

Reruns are not merely hypothetical. In the long thread on model mount errors, Kaggle engineer Dustin wrote on October 24 that earlier rescores were running and that he would see whether submissions hit by the remaining bad cache entries could be included afterward. Later that morning he described a fresh problem in a different region. A repair in one place was not an all-clear.

The public replies reviewed do not say which runs qualify, over what window, how soon results will arrive, or how a competitor will recognize a rerun. CPMP, who started that thread by reporting a model failing to mount before loading began, asked for a replacement submission and later said his team had gone forty-eight hours unable to submit. These are participants' own reports; the desk has not inspected Kaggle's private execution records.

CPMP — Model mount error. · Tanapat Riewruja — Submissions Fails

Zero scores are only part of the question

In the central thread on zero scores caused by dataset loading, Van-Phuc Huynh argued that runs which limped to a positive but depressed score deserve consideration alongside the zeros, and Son Pham replied overnight that the point was fair and that he was unsure of the best remedy. A score above zero does not prove a healthy run, and a low score does not prove a damaged one. Sergei Makarov added reports of early zeros and asked about replacement attempts.

Derrick Goh wants basic execution facts, such as whether the agent initialized and why the run ended, so a broken attempt can be told from a genuine loss. AquaUrsa Research reported checkpoint reads growing from about five minutes to thirty-five with the same configuration, with plain file reads slow before model software was imported, while saying it could not separate the network from the storage backend. Pathik Patel reported that allowing longer for his model server to start recovered a run from zero to about twenty-nine. Those are attributed participant reports, not desk reproductions, and no one has shown that most teams were affected. The central thread's votes show attention, not a count of harmed teams.

Ravi Ramakrishnan — Scoring 0 possibly due to dataset loading issues · Derrick Goh — 8.3 hours, 46.90 on practice games, then 0.00 in under an hour on submission. How do we debug this? · Pathik Patel — Healthy runs scoring 0.00 this week? It's the server-startup deadline (one-line fix)

The deadline question

The last verified final deadline is November 2 at 23:59 UTC, and the rules let organizers change it. A focused check of the staff threads, the timeline page and organizer feeds on the morning of October 27 found no extension decision, no remedy announced by ARC Prize, and no change from the staff's Monday message. Nobody has been contacted for comment by this desk. That is the limit of what the reviewed sources show, not proof that no statement exists elsewhere.

Ravi Ramakrishnan — Scoring 0 possibly due to dataset loading issues · CPMP — Model mount error.

Chen extends the lead to 75.56

On the board itself, Yi-Chia Chen of Yi-Chia Chen @ Fields Model has pulled further clear. A new best score of 75.56, recorded in the early hours, is 4.08 better than the previous 71.48 and holds first place. Tufa Labs remains second at 68.91, a gap of 6.65, and neither its score nor its place has changed in the latest comparison.

Behind them the order at the top is steady. Team Turducken is third at 56.02, the last dance fourth at 50.69, mtg fifth at 50.37 and NVARC3 sixth at 48.84. None of the leading ten moved on the overnight comparison, and Team Turducken, the entry that includes the paper's editor, gets the same scrutiny as every other team. The leading figures are the board's public, provisional standings.

ARC-3 public leaderboard and observed history · Yi-Chia Chen adds 4.08 to ARC-3 lead · Public professional account profile

Climbers and the gold line

The most dramatic overnight move came lower down. bil022 jumped from 0.21 and 3,486th place to 35.26 and 55th, the largest leap in the recorded changes. pensukesan rose from 54th to 27th at 37.22, and Tagir Analyzes climbed from 1,356th to 109th at 33.57. Earlier, the wire noted zhangzibo54 at 37.96 and 23rd.

The provisional gold line was drawn at 38.20 in the 19th place when fshindo returned to the gold places, according to the wire's overnight check. The ARChitects, the four-researcher entry of Daniel Franzen, Jan Disselhoff, David Hartmann and Sebastian Brodehl, stand 16th at 38.85, above that line, and Lord Han Solo is 30th at 37.07. All these medal positions are provisional until the contest closes.

This edition did not refresh social or discussion research; its public archive and notebook are as last saved, and the Kaggle reporting above is dated to the October 27 morning check.

ARC-3 public leaderboard and observed history · fshindo returns to ARC-3 gold places · zhangzibo54 climbs to 23rd on ARC-3 · The 2026 ARChitects: the dream team comes together

From the box score

Team	Rank	Score	Rank change	Score change
Yi-Chia Chen @ Fields Model	1	75.56	Not available	Not available
Tufa Labs	2	68.91	Unchanged	0
Team Turducken	3	56.02	Unchanged	0
the last dance 🕺	4	50.69	Unchanged	0
mtg	5	50.37	Unchanged	0
NVARC3	6	48.84	Unchanged	0
Grigoriy Zatravkin & co	7	47.19	Unchanged	0
復活の混テキスト	8	46.82	Unchanged	0
artificialagencylab.com	9	45.41	Unchanged	0
Majkel1337	10	45.03	Unchanged	0
The ARChitects	16	38.85	Unchanged	0
Lord Han Solo	30	37.07	Down 1	0

Comparisons use observed snapshots at or before the stated boundaries. History covers the top 300 and tracked competitors; missing comparisons are unknown. Score-event entries are context, not net daily gains. Unfamiliar names are not necessarily new entrants.

Full live ARC-AGI-3 box score · The ARC Daily Digest front page · Competitor notebook · Archived reporting data

Kaggle admits run problems. Who gives teams their lost week back?

Kaggle admits run problems. Who gives teams their lost week back?

What is known about reruns, and what is not

Zero scores are only part of the question

The deadline question

Chen extends the lead to 75.56

Climbers and the gold line

From the box score

Sources