LLM-as-Judge

LLM-as-judge is the practice of having a model evaluate output — its own, another model's, or a system's — against stated criteria. It exists because many qualities worth measuring have no assertion: whether copy answers the brief, whether a summary is faithful, whether an explanation is coherent.

It works best under a few constraints.

The rubric should be explicit and itemized, so a verdict is per-claim rather than a global impression, and so a change in the rubric can be tracked as a change.

Code should claim what code can prove. A judge should only be asked about the parts that resist mechanical checking. There is a sharp trap here: an assertion that lives outside the claim it supports cannot fail inside it, so on a broken run the claim is recorded as unproven and silently handed to the judge — buying a model's opinion on something the test had already disproved. The load-bearing assertion must sit inside the claim.

The judge deserves a capable model. Cheaper judges produce notes citing specifics the artifact contradicts, and a note that misquotes its subject is worse than no note.

Judgments are only valid against the inputs that produced them. Change the artifact, the model, the rubric wording, or the criteria set, and the verdict is stale. See Fingerprint and Provenance.