Static site for the TLAPS Proof Benchmark.
The published leaderboard is Proof Completion Core only: every model is graded on the same task list in results/core-manifest.json (currently 190 proof-completion tasks). Proof-from-scratch and Full-suite runs are not shown.
A bundle is published only when it covers that Core set exactly. Older Full-suite dumps that omit Apalache (ben-or83, tendermint) or other Core tasks cannot appear on the leaderboard, even if they were produced this week — their numbers would not be comparable.
The leaderboard's primary score is the Spec-balanced pass rate: calculate the task pass rate within each Core specification, then average those rates so every specification has equal weight. The table also reports tasks passed and specifications completed as supporting counts.
python3 -m http.server 8000Then open http://localhost:8000
- Model runs live in
results/<backend-id>.json(must cover the Core task set exactly). - The Core task list lives in
results/core-manifest.json. - Page copy lives in
scripts/site-content.mjs.
Install the build dependencies once:
npm installAfter changing results, content, or JSX, rebuild:
npm run buildUse node scripts/build-data.mjs --check to validate the generated data without rewriting data.js.
When the Core task set changes:
- Drop a complete Core result bundle into
results/. - Regenerate the manifest:
node scripts/sync-core-manifest.mjs results/<that-bundle>.json- Replace every published model run so each covers the new Core exactly (same benchmarks, theorems, and sources).
- Rebuild and bump the
?v=cache-buster ondata.jsinindex.html.
- Drop the run into
results/<backend-id>.jsonwithmeta.backendset to that id,meta.cohortofone-shotoragentic, currentmeta.scoring, complete usage and equivalent-price data, and results for every Core task. - Add a
BACKEND_INFOentry inscripts/build-data.mjsand list the id inPUBLISHED_BACKENDS. - Add the model's public pricing source to
PRICE_SOURCE_BY_MODEL. - Rebuild, then bump the
?v=cache-buster ondata.jsinindex.html.