Refuse, Decompose, Refresh: A Claim-Safe Protocol for Closed-Loop AI Evaluation

Proposes a protocol to make AI evaluation claims more reliable by refusing, decomposing, and refreshing claims in closed-loop settings.

Auto-scored: A real but partial methodological step toward safer AI evaluation practices, with no direct risk escalation.