AI coding tools can scale mistakes as fast as they scale productivity — generating code that compiles and looks plausible but violates repo conventions, uses deprecated APIs, or misses security baselines. CLAUDE.md and similar instruction files are inputs to generation, not verification systems. The proposed solution is a two-layer verification model: user criteria (per-PR acceptance criteria) and invariants (org-wide rules that auto-apply to every matching change). Invariants form an 'anti-AI slop registry' — codified from recurring review comments — that a separate verifier agent checks against every PR output. An experiment at Aviator generated ~6,000 lines of code from a pre-agreed spec, then verified it against 65 criteria in six minutes, with 60 passing. The model shifts human judgment upstream to spec and intent review rather than line-by-line diff reading.
