AI systems are probabilistic in how they reason. They do not need to be probabilistic in how they act.
Execution authority is a system design choice. When stochastic reasoning is allowed to directly mutate infrastructure, the system no longer has bounded exposure. It has unbounded exposure.
The distinction matters for anyone evaluating risk: the question is not whether a model is unpredictable. The question is whether its authority is.
Governance does not constrain cognition. It constrains the delegation of irreversible authority to stochastic processes. The model remains free to reason. It is not free to act without deterministic authorization.
Reference invariants: ZTG-0b (Observability), ZTG-1 (Replayable Authorization).