Load the spec-writing standard first. Read the complete spec and its referenced documents before reporting anything. This is the pass models do reliably and humans do badly: specs die of internal contradiction more than of missing detail.
Passes
Run all of them; report findings per pass.
- Consistency. Check terminology and every repeated rule across prose, tables, examples, and pseudocode. Any concept with two names, or rule stated twice with drift, is a defect.
- Claim traceability — both directions. Every definition-of-done claim must trace to specified behavior, and every specified behavior must serve some claim or stated goal. Behavior serving nothing is speculative scope; flag it against the exclusions.
- Seam audit. For each cell: responsibility, boundary, contract, and budget all present; contract expressible in the notation standard; the Parnas test holds across each seam. Flag any pair of cells that would change for the same reason (merge candidates) and any cell with two unrelated reasons to change (split candidates).
- Claim classification. Verify each claim's cell-local vs. integration label and the cells it routes through. If most claims are integration-only, report the decomposition itself as the defect.
- Derivability spot-check. Sample decisions the spec does answer: would two capable readers have derived the same answer without it? If yes, it is span — recommend elision. Then sample what it does not answer: would two readers diverge somewhere that matters at a seam, claim, budget, or exclusion? If yes, it is a spanning failure — recommend the missing basis vector or a Deferred Decisions entry.
- Pump hygiene. Confirm amendments since the last review arrived via spec-defect escalation or the catalytic converter, and that converter additions are behavioral claims, not opinions or rationale prose.
Reporting
Report defects with location, the pass that caught each, and a proposed amendment. Do not silently fix: amendments go through HITL. For material specs, offer an independent pass via /request-adversarial-review; use /consensus if findings need adjudication.