AUDITPLAN introduces plan-then-answer framework for auditable safety alignment

AUDITPLAN uses a single-model approach where the agent generates a structured safety plan containing threat labels and constraints before answering. Using the FAITHGATE reward-gating objective, the method improves robustness against unfaithful safety rationales and blanket refusals on Qwen backbones.