Independent LLM-as-judge: scores a candidate output against a task and criteria (0-100), with pass/fail, issues and suggestions. Use it for self-verification, best-of-N, or QA gates.