Closed Form：把 AI 解出的数学题讲给所有人听

Closed Form ∎ — AI mathematics, explained

Live site: https://nishachay.github.io/closed-form/

Every maths problem AI has solved, explained for everyone, with honest trust labels. Each entry starts from a concrete example, says what was known before, names the obstacle the paper removed, and is plain about how certain the result is.

The site is built from the openai/math release (Vol. 1: 372 result families, 722 papers). Explanations are written by an AI pipeline and checked by a second AI pass against the source papers. They are not peer-reviewed. Found a mistake? Every entry has a "Report an error" link. CLOSED_FORM.md is the full spec.

Quickstart

git clone https://github.com/nishachay/closed-form.git
cd closed-form
git submodule update --init --depth 1   # openai/math source, ~2.4 GB
corepack enable
pnpm install
pnpm test     # vitest unit tests
pnpm build    # static site in dist/
pnpm preview  # serve the build locally

Scripts

Command	What it does
`pnpm ingest`	Deterministic rebuild of `src/data/catalog.json` from the submodule (no AI).
`pnpm explain --ids 007,017 \| --batch 10 \| --traces \| --all [--concurrency 20] [--dry-run]`	Reader → checker pipeline; writes `src/content/entries/NNN.json`. `--dry-run` is offline, keyless and write-free. Never run `--all` without a dry run first.
`pnpm calibrate [--apply]`	Re-ranks significance toward the 5/20/45/30 spread. Prints a diff; `--apply` refuses below 10 entries.
`pnpm lint-ste [--entry <path>]`	Sentence-length + hype-word lint over entries.
`pnpm typecheck`	`astro check`.
`npx playwright test`	End-to-end suite in `e2e/` (needs a build first; config serves `pnpm preview`).

Pipeline credentials

Entries are now written by hand from the paper sources and opened as pull requests. The old scripts/explain.ts still reads LLM_BASE_URL, LLM_API_KEY, READER_MODEL and CHECK_MODEL if you run it locally, but no scheduled CI job runs it. Never commit a key. No key is needed for ingest, tests or the site build.

Deploy

Push to main deploys to GitHub Pages via .github/workflows/deploy.yml (actions/deploy-pages, base /closed-form). New entries arrive as PRs, and merging publishes them.

Public data

/data/catalog.json — all 372 families with papers, subjects, trust.
/data/entries.json — explained entries in full.
/data/fields.json — the 17 fields with counts and trust split.
/feed.xml — newly explained entries, newest first.

Contributing

Corrections, bug reports and pull requests are welcome. Read CONTRIBUTING.md and the Code of Conduct. Report security problems privately, as described in SECURITY.md.

License

Code is MIT. Written explanations are CC BY 4.0. Source papers, Lean proofs and traces belong to openai/math under Apache-2.0. Not affiliated with OpenAI.