No download · runs in your browser

Prove It Here

This page runs two proofs, 999 cards total, using the exact same answer cards and the exact same code the downloadable Conduit runs, inside your own browser tab. Nothing is installed. Nothing is sent to a model. THE 300 is the recall proof: 300 questions, with cards built for those questions, replayed exact. THE 700 is the blind-mint proof: 699 cards answered blind, hashed before the key was opened, replayed with zero model calls; 628 of the 699 were correct, clean-blind (89.8 percent), scored once against the public key (tally). When it finishes, your browser computes one hash per proof and shows each one next to the hash we published, so you can see for yourself whether they match. The two proofs are never mixed into one number.

This page is not looking up a saved answer for each question. Each card is a method, a small, deterministic set of steps, and your browser works out the answer from those steps itself, every time you press Run. That is why a card lands on the same right answer every time, and why it keeps landing on the right answer even if the facts underneath it change later.

Methods, not stored answers.

By pressing Run you're using the free replay under the same license as the downloadable Conduit -- see the license. Nothing you do here is uploaded anywhere; the only network traffic is your own browser fetching the public answer cards from this site.

With Cost+20% $5 a year funds about 1,300 proofs for people who come after you. One Cooperative, any doorway: join for $5 a year.

Just want to read one card instead of running all 999? Read One Card.

First Corrected Missing Minted Key notethe answer key is disputed: see the third column

When the answer key is wrong: the third column

Every card is scored against the MMLU-Pro answer key. On a few items the key itself is wrong, or the item has more than one right answer. For those cards we add a third column beside the card: the key's answer, the answer we can support, why, and the source. The notes live in their own file. They do not change any card, any card hash, or either proof hash, and they do not change any score: a noted item is never counted as solved because of its note.

Tier 1Independently confirmed: the MMLU-Redux 2.0 annotators marked the key wrong and named the corrected answer.Tier 2Our own evidence: the computation or cited source shown. No independent correction row exists.

Loading the notes...

Notes file: /prove/data/key_notes.json, published sha256 d238a0ce159e3aba3fb03ebcbc7f8413ac62ccaf1f9f01a695274493e5fd882d. Each note is keyed by the card's qid and the card's sha256 in the MANIFEST. MMLU-Redux 2.0 (dataset, paper) reviewed up to 100 rows per subject of the original MMLU only, so most MMLU-Pro items have no independent row to check. These notes cover the items our records support; a card without a note has not been cleared of key problems.