The Same Action, Graded Twice

The same AI action can be granted in full early in a run and cut to a fraction later, with the request identical both times. A runtime governor sets how much authority an action carries from what the action would do at the actuator and from how far the run has diverged from the trajectory it was assured to follow, and that second reading is not in the action. The runtime governor Wayfinder Systems Group builds sets that grant in the path the instant before the action runs, so what it grants depends on where the run has gone, not on the request alone.

Can the same action be allowed once and refused later?

Yes, and the case is ordinary rather than exotic. An agent proposes an action, the governor grants it, the action runs. Later in the same run the agent proposes the identical action again, the same operation on the same target at the same size, and this time the governor grants a fraction of what it granted before. The request did not change. The run did.

The grant fell because the quantity it is computed from fell. A governor grades an action against where the run is heading, and between the first proposal and the second the run drifted from the trajectory it was assured to follow. The action is the same object both times. The run that action sits inside is not, and the authority a safe step can carry is smaller once the run has turned.

What does the grant depend on that the action does not contain?

The grant is the authority a governor allows a single action, how much that action may do, computed at the actuator the instant before it runs, against a bound declared before the run. A governor computes it from two readings. One is what the action would do at the actuator, in the units the actuator acts on, which is a property of the action. The other is how far the run has diverged from the trajectory the system was assured to follow, which is a property of the run.

The grant is indexed to the run, not to the action. Two requests can be identical down to the byte and still deserve different authority, because the run they sit in moved between them. The governor grades the action as it stands in the run, so the second input, the one carrying where the run has gone, is the input that moves the number. A control that reads only the request in front of it has discarded the one thing that changed.

Why can't you just list the actions an agent is allowed to take?

Because a list fixes authority to the action, and the quantity a safe run turns on is not in the action. An allowlist says this agent may call this tool on this resource, drawn once before the run. That is a permission, and a permission can be precomputed, because what it bounds, the class of thing the agent may touch, does not move while the run executes.

Authority is a different quantity, and it cannot be precomputed the same way. How much a given action may do depends on where the run has gone by the time the action is proposed, and the run's position is not knowable before the run. A bound that holds against a drifting run has to be graded at the moment the action is proposed, against the run as it actually stands. A list drawn in advance cannot carry that, because the thing it would need to read had not happened yet when the list was written.

The two controls range over different objects. A list ranges over actions, the specific calls the agent may make. A governor's envelope ranges over effects, what an action may do at the actuator, and the governor grades each action's effect against that bound given where the run is. A list can be complete and still clear the action that was safe to permit and unsafe to grant in full once the run had turned.

Higher trust, less authority, on the same action.

The sharpest version of the two gradings is the one where the later grant is smaller even though the agent looks more trustworthy than it did at the first. Present-tense trust can read higher on the second proposal, the recent steps clean and the signals calm, and the grant still falls.

A governor can grant less authority on an action whose present trust reads higher, because its forward look caught the run turning before the step landed. The action is identical, the trust is higher, and the authority is lower, because the governor is grading where the run is going rather than how clean the last step looked. When the forward look is uncertain it grants less rather than more, so the identical action late in a drifting run does not inherit the authority it carried when the run was clean.

What the rules ask, and what they leave out.

SR 26-2 took effect April 17, 2026 with generative and agentic AI placed outside its scope and the interagency request for information it promised still not issued. EU AI Act Article 14 asks that a high-risk system be designed so the people assigned oversight can monitor, override, or interrupt it, with high-risk obligations applying from December 2, 2027 for standalone systems and August 2, 2028 for AI embedded as a safety component under Regulation (EU) 2026/1744. The NAIC exposed version 5.0 of its AI Risk Evaluation Supplement, with a later version targeted for the Fall National Meeting in November.

Each asks a deployer to describe the controls it runs. None asks whether a control sets an action's authority from a list fixed before the run or from where the run has actually gone. An action cleared by a list drawn in advance and an action granted against the run as it stands can be the same request and carry very different authority, and only the second reading can tell them apart.

What we are building.

Wayfinder Systems Group builds a runtime governor. It sits above control and below intelligence. It observes, assesses, modulates, enforces, and signs the record. The load-bearing stages are the two in the middle. It sets how much authority each action is granted against a declared envelope and from where the run has gone, and it enforces that grant from inside the path the action takes to the actuator, so the same request gets the authority the run warrants at the moment it is made and no more. We call her Velma. The signed record she emits is the evidence of what each action was granted and why, read after the fact by whoever has to confirm it.