In high-consequence systems, authorization precedes commit.
Financial transactions require pre-authorization. Infrastructure changes require approval gates. Legal instruments require signatures before execution.
AI systems are entering these domains. The ordering constraint does not disappear because inference is probabilistic.
The model may reason stochastically. The authority layer cannot. Governance is deterministic and replayable for authority decisions. This is not a philosophical position. It is a structural requirement inherited from every domain where irreversible execution carries consequence.
Systems that do not enforce this ordering expose the gap between cognition and action to unbounded risk.
Reference invariants: ZTG-1 (Replayable Authorization).