Every graded write-up and audit report behind the accompanying paper, so that the grading can be checked and regraded independently. The AutoFyn runs additionally carry their process trails; the web and provider-agent harnesses expose no trail, so those cells are write-ups and audits only.
Runs are grouped by harness, then model, then run. Cells that were repeated three times have run-01/, run-02/, run-03/ subdirectories; cells that ran once do not.
autofyn/<model>/[run-N/]problem-K/ multi-agent harness runs
claude-code/<model>/[run-N/] provider-agent runs (Claude Code)
codex/<model>/ provider-agent runs (Codex CLI)
zcode/<model>/[run-N/] provider-agent runs (Zcode)
webchat/<model>/ web interface runs
Every run directory holds one audit.md. Each records a frontier model's audit of that run, which a panel of past IMO medalists then reviewed, so the scores in it are the outcome of both stages rather than the model's verdict alone. No model audited its own run: Claude and GLM runs were audited by GPT-5.6, and GPT-5.6 runs by Claude Fable 5.
AutoFyn runs have one directory per problem:
| path | contents |
|---|---|
problem-K/logs.jsonl |
the run's logs, one JSON object per line |
problem-K/current.md |
the working proof state, including the final write-up and its status |
problem-K/approaches/ |
the distinct approaches the run explored |
problem-K/lemmas/ |
lemmas stated and proved along the way |
problem-K/scratch/ |
per-round subagent working notes |
Only logs.jsonl is present in every AutoFyn problem directory. The rest reflect what a given run actually produced, so a run that explored one approach has no approaches/, and four directories carry the write-up as imo-2026-0K*.md rather than current.md.
Provider-agent and web runs expose no tool-call trail, so those directories hold the graded write-up for each problem as problem-K.md, and nothing else for the web runs. Some Claude Code and Zcode runs also kept a code/ directory holding the verification scripts that run wrote and ran. The exception to the naming is claude-code/claude-fable-5/, which uses a problem-K/ directory per problem and carries its own README describing it.
corpus/ holds the fixed pre-2026 corpus the AutoFyn runs retrieve from. past_problems_database.json has 1026 problems from olympiads and national team selection tests between 2006 and 2025, and past_crux_moves_database.json has 2434 crux moves extracted from their solutions. Every entry predates the 2026 contest. The web and provider-agent harnesses carry no corpus.
Every run was audited by a frontier model other than the one that produced it, under a fixed reviewer prompt, and a panel of past IMO medalists then reviewed those audits. Each run's audit file sits beside the write-ups it scored, so both can be read together.
Note that the audit files for the two web runs with unreturned problems report a total over gradable problems only (28/28, 14/14). The paper scores an unreturned problem as zero out of 42, so the same runs appear there as 28/42 and 14/42.