fable-review — try to break it before it ships, don't just nod at it

A review that sets out to confirm an artifact finds little; a review that sets out to break it finds what matters. This skill is the adversarial pass — for code diffs, specs, migration/implementation plans, prompts, and designs — before a decision is locked in. It rates defects by severity; it does not fix them in place (fixing is a separate, explicit step so the reviewer and the fixer stay independent).

Honest scope note: on enumerable, planted defects a single strong model is already near ceiling, so a second independent reviewer added 0 extra recall there (eval/xverify-value/RESULTS.md). The value of an adversarial pass is on judgment/design misses and untested failure paths, not on re-counting bugs a careful first read already caught.

When to use this

• Before merge / ship / finalize / lock-in of a change or artifact.
• The user says "review", "poke holes", "what breaks", "find weaknesses", "red-team it", "what did I miss", "attack this".

When NOT to use this

• You are being asked to implement or fix, not to assess — do that instead.
• A throwaway draft with no decision riding on it.

Procedure

1. Assume it is wrong and find out how. For each part, ask: what input, ordering, failure, or adversarial user makes this break or mislead?
2. Walk the real failure paths: silent failures and fire-and-forget, race conditions, missing timeouts/retries, injection/encoding bypasses, unhandled error branches, boundary and empty/null cases, permission and trust assumptions, claims in a spec that the implementation cannot guarantee.
3. Cite evidence for each finding — the line, the input, the sequence that triggers it. No hand-waving.
4. Rate severity Critical / High / Medium / Low, and say what is needed to close each.
5. Do not rewrite the artifact. Report the findings; let the fix be a separate, deliberate step. If you find nothing real, say so plainly rather than inventing nits.

Expected output

A severity-rated findings list, each with concrete evidence and a close condition — or an explicit "no real defects found; here is what I checked".

Verification

For each finding, could you actually reproduce or point to it? Drop anything you cannot ground — a plausible-but-unverifiable finding is noise. Confirm you reviewed the high-risk surfaces, not just the easy ones.

Failure cases

• Confirming "looks good" without trying any breaking input.
• Listing style nits while missing a real failure path.
• Quietly rewriting the code instead of reporting the defect (collapses reviewer and fixer into one).