Evaluate LLM outputs for factual correctness, hallucinations, bias, uncertainty, policy adherence, and rubric-based scoring.
Evaluation capability embedded inside an assistant rather than run as a separate governance function — checking its own output for groundedness and consistency before returning it.
Interactive assistants where a wrong answer reaches a user immediately and there is no downstream review to catch it.
Batch generation with a review step, where centralised evaluation is more consistent and cheaper than per-response checking.
Self-evaluation must not be the only evaluation. A model checking its own output shares its own blind spots, and an independent evaluation path is still required.
Self-assessment correlating with confidence rather than correctness, so the outputs it rates highest are the ones stated most assertively.
An assistant that declines to answer and says so when its own groundedness check fails, rather than returning a hedged answer that reads like a real one.