Authority Only Falls

An agent's authority is not fixed for the length of a run. It is a value the governor sets before each action, and the direction that value is allowed to move inside the run is the whole control.

Most runtime gates treat authority as a state the agent can return to. That is the gap.

Authority is a value, and the value moves.

A discrete gate answers one question about an action. Allow it, block it, or route it to a person. An earlier piece made the case that the authority an action carried is not one of those buckets. It is a value on a continuum, and the record has to hold the value.

Once authority is a value, a second question follows. The value is not set once at the start of the run and held. It is set again before each action, against what the run has done so far. So the value moves while the agent works. The control is which way it is allowed to move.

Divergence is the signal.

What lowers authority is not an external alarm. It is the distance between what the agent said it would do and what it did. The plan it committed to at the start of a step, against the tool calls it actually made. The state it predicted, against the state it produced.

When realized behavior departs from the assured trajectory, the basis for the authority the agent was granted has weakened. The grant rested on an expectation. The run is now outside it. The signal that lowers authority is the agent's own divergence from its plan, not a detector watching from the side.

The ratchet turns one way.

Inside the run, authority can fall and it cannot rise. Once divergence has reduced it, the agent cannot earn it back by finishing the step. A later action that looks compliant does not restore the authority an earlier divergence removed. Within a single run, authority is non-increasing.

This is an invariant, not a tuning knob. A gate that shrinks scope on a risk score and then re-widens it when the score recovers is not holding this line. The value the governor sets for the next action is bounded by the lowest it has held since the run began.

Why the direction is the control.

A gate that lets authority climb back on the agent's own signal is negotiating with the thing it governs. The agent that diverged is the same agent reporting that it is back on plan. A control that raises authority on that report shares a failure mode with the behavior it is supposed to bound. When the agent is wrong about being back on plan, the control is wrong with it.

The one-way ratchet closes that mode by construction. There is no path inside the run for the agent to talk its authority up. The worst the run has done sets the ceiling for what it is allowed to do next. An examiner can read that ceiling and the divergence that set it.

Recovery is a separate authorized act.

Authority does restore. Not inside the run, and not on the agent's say-so. It restores through a fresh authorization. That authorization is recorded, held by a party under its own envelope, and bound to the specific scope being restored. It is a new admission, not a continuation of the old grant.

So the record has to hold three things. The reduction, the divergence that caused it, and the fact that any restoration was a separate act by an authority that was allowed to grant it. A run that quietly returns to full authority after a divergence, with nothing on the record marking the reduction or the grant that reversed it, cannot show an examiner that the control was ever in force.

Why an examiner asks now.

SR 26-2 took effect on April 17, 2026 and placed agentic AI outside its scope, pointing institutions back at their own risk practices while an interagency RFI on AI is still forthcoming. EU high-risk obligations moved to December 2, 2027 for standalone systems and August 2, 2028 for embedded ones. The NAIC's AI evaluation tool is in a multistate pilot running into the fall, giving examiners a standardized way to review how an insurer governs its AI.

None of those dates changes what the examiner does at the end. The examiner pulls one action that affected a person and asks what governed it. If a divergence lowered the agent's authority and the run quietly handed it back, the record has to be able to say so. When it cannot, the control was not in force, whatever the policy said.

What we are building.

Wayfinder Systems Group sets the authority of each agent action before it runs, lowers it when realized behavior diverges from the assured trajectory, and holds it non-increasing for the rest of the run unless a separate authority restores it on the record. The reduction, its cause, and any restoration are signed onto a tamper-evident chain at the moment each one happens. Patents held in The Wayfinder Trust. We call her Velma.