Iterates on a failing AI/LLM eval (an L2 suite, a golden-set / judge eval, or any eval gating a PR) until it is green AND confirmed, not just luckily passing once. Classifies the failure as a code bug, an eval-definition bug (stale golden item / criteria drift), judge-drift (silent grader version change), or flaky. Applies the minimal fix, then requires N consecutive confirming re-runs — 2 deterministic, 5 for anything a model call grades, since a flat "2" is statistically weak for a stochastic judge. Hard-capped at 5 fix iterations. Refuses to game the eval (Goodhart's Law: skip/delete/overwrite-in-place a case, loosen a threshold) without a second independent check plus `confidence(analysis) >= 90%` and a logged rationale. Composes `ai-engineering`, `confidence`, `critical`, `verify-behavior`. Use when an eval is failing on a PR and needs a real, non-gamed green. Triggers on "this eval is failing", "iterate on this eval", "fix this eval", "get this eval green", "optimize this eval", "/eval-iterate".