The Point of No Return

An agent action does not become real when the model decides it. It becomes real when the effect lands. A payment leaves an account. A row is written to a system of record. A message reaches a customer. Before that moment the action is still contingent, and it can be reduced, held, or refused. After it, there is only the account of what already happened.

The examiner pulls that one action and asks what governed it. A control that read the action afterward can describe the effect. It could not change it.

The moment the effect becomes durable.

Every agent action has a boundary. Before the boundary the effect is contingent and still inside the run's control. After it the effect is committed to the world and outside it. The boundary is a single step. Usually it is the tool call that writes, sends, moves, or actuates.

A proposal is contingent. A tool call that debits an account is committed. The interesting question about any governance control is not how good its verdict is. It is which side of that step the control sits on when it renders the verdict.

Two ways to be on the wrong side.

The controls most firms hold sit on the wrong side of the boundary, in one of two directions.

A control set before the run is too early. It was fixed at provisioning, before the action existed, and it cannot read the action it never saw. That is the argument of the companion piece. A clean pre-deployment evaluation does not reach the decision made at inference.

A control that reads the action after the effect landed is too late. It can produce an account, and the account can be detailed and honest. It is an account of an effect already in the world. The only control that can govern the action is the one that sits in the decision path, before the effect becomes durable.

Why after is only a description.

Consider what a firm typically has once an agent has acted. An explanation. A log. An alert. A drift signal. Each of these is produced at or after the moment the action ran.

Each one describes. None of them decided how much of the action's effect was allowed to land, because by the time any of them exists the effect has already landed. A description of an irreversible act is evidence about the past. It is not authority over the act. Authority is exercised on the near side of the boundary or it is not exercised at all.

The record has to close on the near side too.

This is the harder consequence, and it is the one most runtime designs miss. If the control has to sit before the durable effect, the record it signs has to close before the durable effect as well.

A record written after the tool call was produced by the same run that produced the effect, which means the party under examination could have shaped it. A record sealed before the effect lands fixes what the governor decided while it still held the authority to decide. The seal has to close before the point of no return, not after it. The examiner needs the near-side record, because that is the one the actor could not revise once the effect was already done.

Where the line is drawn 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. Enforcement layers are shipping across the agentic stack that allow or block a tool call at sub-millisecond latency. Those layers sit on the near side of the boundary, which is the correct side. Most of them do not seal a record there.

EU AI Act high-risk obligations were deferred to December 2, 2027 for standalone systems and August 2, 2028 for systems embedded in regulated products. The runway is real. The boundary does not move because the deadline did. When the examination cycle arrives, the question is not whether the firm noticed the action. It is whether a control stood on the near side of the point of no return and left a record sealed there.

What we are building.

Wayfinder Systems Group sits in the decision path, on the near side of the durable effect. It reads each action the agent proposes, sets the authority of that action while the action is still contingent, and signs the value onto a tamper-evident chain before the effect can land. The enforcement decision and the record close on the same side of the point of no return. Patents held in The Wayfinder Trust. We call her Velma.