Escaping the Agreement Trap: Defensibility Signals for Evaluating Rule-Governed AI

TL;DR

This paper introduces a new framework for evaluating rule-governed AI systems by focusing on policy-grounded correctness and defensibility signals, addressing limitations of agreement-based metrics.

Contribution

It formalizes policy-grounded evaluation metrics, introduces the Defensibility Index and Probabilistic Defensibility Signal, and validates the approach on Reddit moderation data.

Findings

Agreement-based metrics underestimate true correctness by 33-46.6 percentage points.

Policy-grounded decisions account for 79.8-80.6% of false negatives.

Auditing under different rule tiers reduces ambiguity by 10.8 percentage points.

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.