From Declared Principles to Executed Policy: Open-Set Auditing of Factorized Moral Functions

An artificial agent may produce decisions consistent with a declared moral policy while implementing a materially different decision process. Standard benchmarks and closed-set preference-elicitation methods can miss this discrepancy because they ask which candidate policy best explains observed outputs, not whether any candidate is adequate. We represent the executed policy through five audit roles: perception, qualification, normative weighting, decision, and control. The factorization is used to design diagnostic interventions and to localize deviations, while behavioral equivalence classes define what can actually be identified from a finite query family. In a reproducible synthetic proof of concept motivated by the multi-attribute structure of lung allocation, 72 policy descriptions yield 48 behavioral equivalence classes. At 5% response noise, generic active selection identifies 89.7% of classes within five queries and 99.7% within ten. Factor-informed interventions do not outperform well-selected natural cases in this closed setting. Their value emerges under misspecification: for two out-of-class mechanisms almost indistinguishable on natural cases, intervention-based auditing raises calibrated rejection from 8.0% to 97.3% and from 4.0% to 98.8%, without increasing in-class false rejection. The central result is therefore not that factorization replaces generic active learning, but that it supports open-set audit: it exposes mechanisms hidden by ordinary cases and separates confidence in the best available explanation from evidence that the explanation is adequate.