opinion
The model should not be the policy engine
Semantic reasoning can inform policy decisions, but hard constraints, permissions, budgets, and operational invariants should remain outside the reasoning model wherever they can be deterministic.
The model may be the best part of a system at interpreting ambiguous intent. That does not make it the right authority for granting itself permission.
Keep hard constraints, permissions, budgets, and operational invariants outside the reasoning model where they can be deterministic. Let semantic judgment inform a policy decision when necessary, but keep the final boundary in a runtime, credential, or infrastructure layer that can be inspected and enforced independently.
This is an architectural hypothesis, not a report of a completed implementation or hands-on test. It connects directly to Authority Compiler and A runtime for agents is an authority system.