Logging confirms that something happened. Governance determines whether it is allowed to happen. These are different architectural functions.
Most current AI risk mitigation relies on monitoring: detect an adverse event, escalate, respond. This model works when actions are reversible and consequences are bounded.
When an AI system executes an irreversible action — commits a transaction, mutates production infrastructure, sends a communication with legal effect — monitoring records the outcome. It does not prevent it.
Governance must exist between cognition and action. Not after action, in the form of alerting. Not alongside action, in the form of logging. Before action, as a mechanistic prerequisite.
Where governance and execution occur within the same temporal sequence, control collapses into observation.
Reference invariants: ZTG-1 (Replayable Authorization), ZTG-4 (Mechanistic Boundary).