arXiv:2604.20972v1 Announce Type: new Abstract: Content moderation systems are typically evaluated by measuring agreement with human labels. In rule-governed environments this assumption fails: multiple decisions may be logically consistent with the governing policy, and agreement metrics penalize valid decisions while mischaracterizing ambiguity as error - a failure mode we term the Agreement Trap. We formalize evaluation as policy-grounded correctness and introduce the Defensibility Index (DI) and Ambiguity Index (AI). To estimate reasoning stability without additional audit passes, we introduce the Probabilistic Defensibility Signal (PDS), derived from audit-model token logprobs. We harness LLM reasoning traces as a governance signal rather than a classification output by deploying the audit model not to decide whether content violates policy, but to verify whether a proposed decision is logically derivable from the governing rule hierarchy. We validate the framework on 193,000+ Red...