fable-review

# 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, injec...