Definition
Plain language
Using a language model to grade an answer that can't be checked by running code, like a written math proof.
As stated in the literature
A learned LLM-based verifier that scores or critiques outputs lacking mechanical checks; in MaxProof, a four-layer stack (rules gate, normalizer, three heterogeneous judges, minimum aggregation) designed to minimize false-positive rate over a long training stream rather than maximize benchmark accuracy.
Why it matters: It lets systems grade answers that can't be checked mechanically, which is essential when correctness lives in prose or reasoning rather than in runnable code.
For example, a language model reads a written math proof and judges whether each step truly follows, since there's no program that can simply run the proof to check it.