What is an eval?

The evaluation endpoints are in development and are not part of the published API yet.

An eval is a judgement, recorded as a run, about whether a piece of content holds up. It asks whether the claims are supported by a source, and whether it matches the brand it is published under.

Two properties make an eval worth keeping. Accuracy and brand are judged as separate calls, so a good score on one cannot mask a bad score on the other. And runs are append-only, each recording the exact text it judged, so a verdict stays meaningful after the underlying source changes.

Verdicts are states rather than scores (supported, failing, unverified) because models are far more consistent at picking between a few classes than at producing a stable number.