Review

Every finding cites evidence — file:line or a source. No evidence → flag [unverified]. Default to refuted: a flaw you cannot prove is a flaw you note, not one you wave through.

An LLM cannot self-correct on its own judgment — left alone it drifts or degrades. Review fixes that the only way that works: a separate skeptic anchored to an external oracle — the code, §R, the test suite, the docs. "Looks good" is not a review. A refutation attempt is.

WHEN TO REVIEW

• Before /build on a high-blast-radius change (shared module, auth, data, money, public API).
• Spec touched §I or §V that other code depends on.
• Right-sizing says the cost of a wrong build > the cost of one review pass.

Skip for a trivial, reversible, well-understood change. Adversarial review on a typo hallucinates flaws & wastes the budget — the self-critique paradox is real.

PHASE 0 — CAPTURE

Read the spec: §G §C §I §R §V §T. Hold the whole thing. You review the spec, not your memory of the conversation.

PHASE 1 — CONSTRUCT THE SENIOR

Build a reviewer with real authority, not a generic critic:

• Codebase — grep/read the modules this spec touches. What patterns, what invariants already hold?
• §R — what did research establish? A spec decision that contradicts §R is a finding.
• Live — for any best-practice claim you are unsure of, fetch it. An out-of-date assumption is a flaw.

A reviewer with no evidence is just an opinion. Earn the authority first.

PHASE 2 — REFUTE

Attack the spec on these axes. For each, try to find the case where it breaks:

• Goal vs reality — does §G solve the actual problem, or a proxy?
• Missing invariant — what can go wrong that no §V catches? (most findings live here)
• Interface drift — does §I match what callers already expect? (cite the caller, file:line)
• Constraint conflict — do two §C bullets contradict? does one fight §R?
• Unowned edge — the input, ordering, failure, or concurrency case no §T covers.
• Altitude — §T too vague to act on, or so granular it is just typing?

PHASE 3 — CLASSIFY

Each finding: evidence → claim → severity.

• BLOCK — build on this spec ships a real defect. Must fix first.
• HARDEN — add/sharpen a §V so the build cannot regress it.
• NOTE — worth knowing, not blocking.

No evidence? Down-rank to NOTE & tag [unverified]. ⊥ inflate a hunch to BLOCK.

PHASE 4 — HARDEN §V & GATE

• Each HARDEN finding → a draft §V line (testable, cites the §I/behavior it guards). Hand to spec to write.
• End on an explicit gate:

## review verdict
BLOCK: 1 — §I.api shape ≠ caller src/client.ts:40. fix §I before build.
HARDEN: 2 — drafted V8 (idempotent refund), V9 (tx around dual write).
NOTE: 1 — §T4 vague, split before /build.
gate: NO-GO until BLOCK cleared. then /build §T after spec writes V8,V9.

GO or NO-GO, never a shrug. Review is the checkpoint that stops a confident wrong build.

BOUNDARIES

• ⊥ write SPEC.md. Draft §V & hand to spec.
• ⊥ pass a finding with no evidence as fact. Flag [unverified].
• ⊥ review trivia. Right-size or skip.
• ⊥ rewrite the user's intent. You harden the spec, you do not replace its goal.