Blog
Runs, results and what didn't work.
Dated Kaggle competition runs with every rejected experiment, and practical guides to evidence-first ML work from Claude Code, Cursor or the command line.
Evidence from ARC-AGI-3, Enveda CASMI 2026, Squares in Squares and Gemma 4
Autoresearch for Kaggle: What 136 Overnight Rounds on Four Live Competitions Taught Us
Karpathy’s autoresearch loop on live Kaggle competitions: what had to change, a medal-range result, 136 recorded rounds and the lessons, failures included.
ARC-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.
ARC Prize 2026 — ARC-AGI-3 · Kaggle · $850,000 prize pool
ARC Prize 2026 (ARC-AGI-3): A Bronze-Medal-Range Run and Every Rejected Idea
ARC Prize 2026 ARC-AGI-3 on Kaggle: 9-hour no-internet runs, closed November 2, 2026, results Dec 4. Our rank: 313 of 4,926, bronze range. Every idea shown.
Enveda CASMI 2026 · Kaggle · $50,000 · closes 14 Dec 2026
Enveda CASMI 2026 on Kaggle: From Rank 3,856 to 1,421, 0.008 From Bronze
Our Enveda CASMI 2026 run: 52 rounds, public score 0.136 to 0.479, rank 1,421 of 5,305 and 0.008 below the bronze-medal line. What worked, what failed.
Squares in Squares · Kaggle Playground · $50,000 · closes 18 Jul 2027
Squares in Squares on Kaggle: 46 Rounds, 22.2502 to 22.1756, and a Moving Field
Our Squares in Squares run: 46 rounds took the public score from 22.2502 to 22.1756, rank 326 of 2,052. Which merges held and which of our ideas did not.
Gemma 4 Developer Agent · Kaggle · $75,000 · closes 2 Dec 2026
Gemma 4 Developer Agent on Kaggle: Noise, No-Op Adapters and Failed Runs
Our Gemma 4 Developer Agent run: 0.17 public, rank 2,249 of 5,691, bronze line 0.20. An identical resubmission moved 0.05; 20 of 24 rounds failed on compute.
Evidence from Enveda CASMI 2026 and ARC-AGI-3
Kaggle CV vs Leaderboard: What to Trust When They Disagree
Your Kaggle CV and public leaderboard disagree: which do you trust? A recorded case where the public score rose 0.136 to 0.166 while local MRR@25 fell.
Evidence from Gemma 4 Developer Agent and ARC-AGI-3
Kaggle Leaderboard Noise: The Same Submission Scored 0.12, Then 0.17
We resubmitted a byte-identical Kaggle submission and the public score moved from 0.12 to 0.17. How to measure leaderboard noise before you trust a gain.
Claude Code vs Cursor for AI Research Workflows: Where Each One Wins
A practical comparison of Claude Code and Cursor for research synthesis, coding execution, and evidence-based iteration.
More posts
The AI Research Benchmark Checklist Every Team Should Use Before Shipping
A practical checklist for baselines, hypotheses, validation, and evidence before calling a result a winner.
Reproducible AI Research Without Notebook Chaos: A Practical Workflow
A repeatable research workflow for teams that want traceable evidence instead of scattered notebooks and vague “wins.”
Video walkthroughs
Each run, explained on the Forger Labs HQ YouTube channel.