Deterministic no-LLM judge for tests and CI (spec §4).
Not a real evaluator. It grades a criterion by a trivial, fully
deterministic rule: the criterion passes iff its description text appears
(case-insensitively) somewhere in the flattened transcript. This makes grading
a pure function of the rendered description, so a description-tuning regression
(a terser variant that drops a required phrase) surfaces as a failing criterion
without any model in the loop.
Real LLM-as-judge adapters are app-layer follow-ups; select one via the
:noizu_mcp :eval_judge application env.