false-accept
Evidence home for the verdict-layer approach: pre-registered, reproducible trust evaluation of AI-generated code changes - silent wrong ships 17/50 → 0/50.
