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

arXiv:2604.20972v1 Announce Type: new Abstract: Content moderation systems are typically evaluated by measuring agreement with human labels.