AUDITPLAN: Commit, Then Answer for Auditable Safety Alignment
Proposes a two-phase protocol that forces models to commit to a plan before answering, enabling external auditors to verify safety alignment.
Proposes a two-phase protocol that forces models to commit to a plan before answering, enabling external auditors to verify safety alignment.
Auto-scored: A real but partial technical step toward auditable safety alignment, with no immediate risk escalation.