🤖 AI Summary
This study addresses the “consistency trap” in traditional evaluation methods, which erroneously penalize rule-compliant decisions that deviate from historical human labels. To overcome this limitation, the authors propose a rule-based correctness evaluation framework that validates model decisions through hierarchical logical reasoning grounded in governance rules. They introduce two novel metrics—the Defeasibility Index (DI) and Ambiguity Index (AI)—to quantify assessment quality. Innovatively leveraging reasoning traces from large language models as governance signals, they formulate a Probabilistic Defeasibility Signal (PDS) to estimate inference stability without manual review. Experiments on over 193,000 Reddit moderation decisions reveal a 33–46.6 percentage-point gap between consistency-based metrics and rule-based correctness, with 79.8–80.6% of false negatives actually being compliant. A governance gating mechanism achieves 78.6% automation coverage while reducing risk by 64.9%.
📝 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+ Reddit moderation decisions across multiple communities and evaluation cohorts, finding a 33-46.6 percentage-point gap between agreement-based and policy-grounded metrics, with 79.8-80.6% of the model's false negatives corresponding to policy-grounded decisions rather than true errors. We further show that measured ambiguity is driven by rule specificity: auditing 37,286 identical decisions under three tiers of the same community rules reduces AI by 10.8 pp while DI remains stable. Repeated-sampling analysis attributes PDS variance primarily to governance ambiguity rather than decoding noise. A Governance Gate built on these signals achieves 78.6% automation coverage with 64.9% risk reduction. Together, these results show that evaluation in rule-governed environments should shift from agreement with historical labels to reasoning-grounded validity under explicit rules.