← All posts
Progress reportARC-AGI-3 · Enveda CASMI 2026 · Squares in Squares · Gemma 4 Developer Agent

Our Kaggle Standings, November 2026: One Bronze-Medal Range, Three Live Runs

Our standings on four Kaggle competitions, 14 Nov 2026: ARC-AGI-3 in the bronze-medal range, Enveda 0.008 from bronze, and what each run taught us.

4
Competitions
ARC-AGI-3
In medal range
Enveda · 0.008
Closest live
136
Rounds on record

Manas Joshi·November 14, 2026·8 min read

We run ResearchForge on four Kaggle competitions in the open. This is where each one stands on 14 November 2026, against Kaggle’s medal lines, and what the record taught us since the last update. Every rank is from the public leaderboard; the 136 rounds behind them are on file, including the 47 that failed or were rejected.

CompetitionRankTopBest publicWhere it stands
ARC-AGI-3313 / 4,9266.4%32.51🥉 Bronze-medal range; silver line 32.97
Enveda CASMI 20261,348 / 5,26125.6%0.4790.008 below the bronze line (0.487)
Squares in Squares326 / 2,05215.9%22.1756Playground, no medals; top 10% is 22.1617
Gemma 4 Developer Agent2,249 / 5,69139.5%0.17Bronze line 0.20

Kaggle awards medals on the final private leaderboard. With 1,000 or more teams, bronze goes to the top 10%, silver to the top 5% and gold to the top 10 plus 0.2% of the field, so these are standings, not medals yet.

ARC-AGI-3: inside the bronze-medal range

Submissions closed on 2 November with us at rank 313 of 4,926, comfortably inside the bronze cut-off at rank 492 and 0.46 short of silver. The final standings use hidden games and arrive on 4 December, and our own repeat submissions ranged from 21.75 to 32.51, so the rank can move either way. The full run, with all 10 rejected ideas, is in the ARC-AGI-3 write-up.

Enveda CASMI 2026: from rank 3,856 to 1,348 in nine days

On 5 November we were at 0.166 and rank 3,856. The overnight loop then ran 34 more rounds. Our own ideas lifted local MRR@25 from 0.889 to 0.915 and the public score to 0.213. Adopting a public pipeline whose author reported 0.500 took us to 0.479 and rank 1,348, 0.008 below the bronze line with a month to go.

📌
What we learned: in October the public score rose while our local score fell. In November they moved together in five of seven steps, but a 0.0008 local gain was worth +0.022 public while other sub-0.001 steps were ties. Follow the local direction; don’t read its size. And a public notebook’s reported score is one draw: ours scored 0.479, not 0.500. Details in CV vs leaderboard.

Ten adopted public pipelines ran to the end without printing a local score, so the record marks them unscored rather than guessing. Only the leaderboard can judge those, and daily submissions are limited, so the loop has to choose which to spend them on. Every round is in the Enveda CASMI 2026 write-up.

Squares in Squares: 46 rounds, and the field keeps moving

Our public score improved from 22.2502 to 22.1756 (lower wins), yet our rank went from 174 to 326, because 400 more teams joined and the whole leaderboard tightened. 17 rounds improved the score and 16 of those were certified merges of public packings, checked for overlap before anything was submitted. Our own search ideas, polishing, neighbour transfer and new seed shapes, ran 12 times for one gain of 0.0000015.

📌
What we learned: in a competition where everyone shares packings, the merge is table stakes and speed matters: the loop now watches public notebooks and merges new outputs as they appear. Its newest local packing measures 22.1580, which would sit inside the 14 Nov top 10%, but it is not submitted yet, so it is not in the table above. One merge that passed our own check came back from Kaggle’s scorer as an error: a local check is not the scorer. The whole run is in the Squares in Squares write-up.

Gemma 4 Developer Agent: compute is the bottleneck

Still 0.17, now rank 2,249 of 5,691, with the bronze line at 0.20. The headline finding is still the byte-identical resubmission that scored 0.12 and then 0.17. Of 24 recorded rounds, 20 failed, most of them on Colab compute: no GPU available, sessions dropping or timing out. The best local benchmark reached 0.40 on 10 tasks and 0.33 on 30, but a failed run is not a measurement, so nothing from that stretch has earned a submission. Making those runs finish is the next job. Details in the Gemma 4 write-up.

What all four runs taught us

  • ▸Our biggest jumps came from strong public work. On ARC-AGI-3, Enveda and Squares alike. ResearchForge’s job was to check it, keep what held up and record what didn’t, and we say so plainly.
  • ▸A local benchmark is only as good as what it holds out. It misled us on ARC-AGI-3 and on Enveda in October, and helped on Enveda in November.
  • ▸One leaderboard number is a sample. 0.12 vs 0.17 on Gemma, 21.75 to 32.51 on ARC, 0.500 vs 0.479 on Enveda.
  • ▸Failures are data only when the reason is kept. 47 of 136 rounds failed or were rejected, each with its reason, so none gets re-tried by accident.

What’s next

ARC-AGI-3 results arrive on 4 December, Gemma 4 closes on 2 December, Enveda on 14 December and Squares in Squares on 18 July 2027. We will update this page and each competition run as they move. To run the same loop on your own competition, start with Kaggle Overnight.

Try it on your next run

Every number in this post came from the ResearchForge record. Yours can too.