Definition
Plain language
A checklist that grades whether an AI's math proof is grounded, faithful, complete, and well-written.
As stated in the literature
In RMA, a verifier module enforcing five rules on candidate proofs: claim grounding, problem faithfulness, no logical gaps, explicit constructions where required, and clean formal notation.
Why it matters: Explicit, rule-based grading of proofs catches the slick-sounding gaps that final-answer scoring would miss entirely.
For example, the module flags a proof that asserts 'such an x exists' without actually constructing it, refusing to mark it complete.
Heard on the show
“The third — and this one is doing real conceptual work — is what they call the Proof Commandment Module.”Episode 076 — Same Model, Organized Differently: How an Agent Architecture Beat Frontier Systems at Research Math