Arcifact Verify
Public claims, receipts, and the verifier for Arcifact Chess.
Arcifact Chess measures playing strength from the public Lichess database: twelve months of rated games, 88.2 million evaluated records. A claim is published only after it passes a pre-registered calibration gate. Failed gates are published too, in the same ledger.
Every published figure comes with a receipt: the data slice, the seed, the engine version, and a result hash. The verifier in this repository recomputes figures from the public data and checks the hash. It runs on your machine and contacts no server of ours.
Contents
| Path | Contents |
|---|---|
claims/GATE_LEDGER.md |
Every gate run to date: passes, failures, and work in flight. |
claims/CORRECTIONS.md |
The corrections register. |
receipts/PUBLIC_SCHEMA.md |
The receipt record format. |
verifier/verify.py |
The verifier. No dependencies beyond Python 3.10. |
Quick start
python3 verifier/verify.py selftest
This runs the bundled fixture end to end in about a second. To verify a corpus receipt against a real Lichess monthly dump (large downloads; a month of blitz is tens of gigabytes):
zstdcat lichess_db_standard_rated_2026-06.pgn.zst | \
python3 verifier/verify.py check receipt.json -
Corpus receipts for the twelve-month dataset are issued with the public launch. Claim-level recomputation recipes follow in the verifier's next release.
Method
Claims start as pre-registrations with sealed hashes. Calibration uses dual null designs: interval coverage is scored against simulated truth, and false-fire rates are measured on constant-truth data. Confidence is stated per tier; 95% is the primary standard, and the realised error we measure is published next to it. Where the data cannot support an answer, the ledger says so.
Relation to existing rating systems
Lichess ratings use Glicko-2, which maintains a rating deviation and volatility internally; that machinery exists to run matchmaking and predict results inside its own pool. Arcifact does not claim to discover rating uncertainty. It provides a separately calibrated, interpretable measurement record: an interval whose coverage we test and publish, change verdicts with measured false-fire rates, personal findings that pass predefined gates, corrections at claim size, and refusal when the evidence is insufficient.
Corrections
If a published figure does not replay for you, open an issue with your recomputation. Substantiated corrections are published in the register at the size of the original claim, with credit.