The happy path is insufficient. The control boundary must be challenged.
AX-01Alternate route
Can the same consequence be reached through an undeclared or secondary execution path?
Proof target: All materially equivalent routes are blocked, separately governed, or explicitly declared out of scope.
AX-02Stale authority token
Can a previously valid authorization be replayed after authority or conditions have changed?
Proof target: The stale token is refused before consequence and the refusal is attributable.
AX-03Renamed or aliased tool
Can a prohibited capability be reached by changing the route, tool name, alias, or wrapper?
Proof target: Control follows consequence and capability identity rather than a single label.
AX-04Direct API bypass
Can an actor skip the governed adapter and invoke the consequential surface directly?
Proof target: Direct access is impossible, separately authorized, or preserved as an explicit limitation.
AX-05Privilege expansion
Can an actor or agent obtain broader execution power after the original approval?
Proof target: Privilege expansion invalidates standing and forces a new bounded determination.
AX-06Race and delayed commit
Can conditions change between final evaluation and irreversible execution?
Proof target: The commit boundary is atomic enough, short-lived enough, or revalidated before consequence.
AX-07Fail-open behavior
What happens when the policy engine, evidence service, network, adapter, or verifier becomes unavailable?
Proof target: Failure behavior is explicit and cannot silently expand permission.
AX-08Dependency substitution
Can a model, dataset, sensor, policy, endpoint, or other material dependency change after approval?
Proof target: Material substitution breaks correspondence and triggers revalidation before execution.
AX-09Out-of-band execution
Can the consequence occur outside the runtime surface observed by the governance system?
Proof target: Out-of-band pathways are controlled or preserved as a bounded non-claim rather than hidden.