Step ④ — Verify "done AND good enough", independently

Why this exists: the builder cannot be the verifier. Their self-check already happened in Step ③; reusing it as the gate is exactly the self-grading failure the whole loop exists to stop. This is not about distrust of a person — it's structural. In the source system's own logs, a detector self-reported 1.00 recall; an independent test measured 0.17. Writing the code, testing it, and declaring the result form one unbroken chain of confirmation bias.

Do

An independent reviewer (≠ the builder) checks each stage against the frozen acceptance table and issues a per-stage verdict — done AND good-enough. All hard gates pass → deliver. Any fail → Step ⑤ (remediation-loop).

The reviewer must have the frozen table in hand — verifying from the builder's self-narration does not count.

The metric self-review ban (the crown jewel)

For any measurable claim — recall, precision, cost, coverage, latency, accuracy — the normal verdict is not enough. The reviewer must:

• Not read the implementation first (review against the criteria / goal, not the code).
• Recompute ≥3 samples by hand.
• Have before/after data points to compare.
• For a claimed metric: cite sample provenance (≥50 independent + ≥10 adversarial), with thresholds justified by an external basis.

See references/hard-gates.md for the full truth table of when each gate fires.

Independence

Reviewer ≠ builder. A self-check from Step ③ can never stand in for Step ④. For metric-validity claims, use the stricter hand-recompute path above; for ordinary high-risk work, an independent reviewer reading the frozen table is enough.

Relationship to superpowers

superpowers requesting-code-review and verification-before-completion give you the mechanics (how to run the review, how to gate completion on evidence). This skill adds what they don't carry: the per-stage verdict against a frozen table and the metric self-review ban.