AI teams can't trust their own tests anymore.

Teams ship new versions of AI agents every week, and they decide what ships by running tests. But an AI system can learn its own test — through training data overlap, through repeated exposure, through the test simply being visible to the pipeline. The score goes up; the ability doesn't. Nobody finds out until production.

Whetstone is the promotion gate: a private, constantly-refreshed exam, checked against every declared training set so the model hasn't seen it, graded by exact checkers instead of answer keys, with a statistical decision — PASS, HOLD, or BLOCK — and an audit trail for why every version shipped or didn't.

This button runs the real machinery right now: it mints fresh exam items against an exact verifier, checks them against a training set and quarantines the overlap, grades two systems with no answer key anywhere, and issues a promotion decision with an append-only ledger. The two systems under exam are stored answer policies — stand-ins so this runs without a GPU — but every grade is computed live by the verifier.

Same private 48-item exam, three versions of the same 4-billion-parameter model, graded on the machine in this room. The middle one scores higher than base — and the gate refuses it, because two reliable code abilities quietly broke. The third routes the fine-tune to where it helps and steps out of the way where it hurts. That one earns promotion.