A registry entry begins with a primary source.

What happens to a submitted paper

1. DOI, preprint ID, canonical URL, title fingerprint, open candidate issues, and recent paper-intake PRs are checked for duplication.
2. The source is MIME-checked, capped at 45 MiB / 150 pages, fingerprinted, and read only from an ignored local temporary directory.
3. After the owner explicitly selects the paper, a fresh read-only Codex session separates evaluation, training/fine-tuning, validation, model selection, external-result summary, and pure background citations.
4. A second fresh Codex session independently rechecks every claim and locator against the original source. Disagreement never becomes a normal field.
5. Deterministic code—not either session—assigns IDs and creates Work, BenchmarkUse, Model, and sufficiently complete EvaluationRun records.
6. A Ready PR runs normal CI and can merge only after wang422003 comments /approve-paper-intake <full-current-head-sha>.

Pull request checklist

• Describe what changed and cite a primary or official source.
• Map Scientific Tasks only from an official taxonomy, formal track, or locatable official artifact; preserve count unit and basis.
• Say whether the evaluation is full, subset, track, or unknown—and include realized n when reported.
• Record metrics, aggregation, tools, repeats, grader, and every remaining unreported field.
• Do not copy benchmark questions, restricted attachments, or third-party papers into this repository.
• Run pnpm registry:validate and pnpm registry:build.

Review policy

All production changes use pull requests. Paper-intake PRs additionally require the paper-owner-gate: only an exact-SHA approval comment by wang422003 passes, and any later push invalidates it. Local Codex never auto-merges. Candidate issues, source files, evidence excerpts, session transcripts, and model drafts are excluded from public exports.