reviewer

mode: subagent — the final validation step, the hard PASS/FAIL gate.

What it does

Calls build_review_evidence to get a pre-computed block of every deterministic check (Knox validation, output shape/schema, Nyquist, plot manifest), then emits REVIEW: PASS or REVIEW: FAIL based solely on those numbers — it never re-derives a check itself. See Determinism & no fabrication for why: it's the one step that must not depend on model quality.

@plot_reviewer is a separate, advisory-only agent — see plot_reviewer — it never gates the run.