Time to Receipt

A governed agent action produces a sealed record. The standing question about that record is how much the sealing slows the agent. That is the wrong question. The record has a deadline, and the run itself sets it.

The throughput frame is incomplete.

The common way to reason about in-path governance is throughput. A governor that fires before the next action sits in the decision path, and the concern is how many microseconds each decision waits. Measured that way, the governor is a tax on speed, and the engineering goal is to make the tax small.

Speed matters. It is not the whole obligation. Framing time to receipt as a throughput cost treats the deadline as something the operator chooses. The operator does not choose it. The structure of the run does.

The deadline the run sets.

An agent action lands an effect. A row is written. A balance moves. A message goes out. A later step in the same run reads that effect and acts on it. The second action now depends on the first.

Between the moment the first effect lands and the moment its record is sealed, there is a window. If a second action depends on the first inside that window, the run has built on an action that is not yet provable. Two actions are entangled, and only one of them is on the chain. The record has to seal before the effect it documents becomes something another action relies on.

That is the deadline. It is not a latency budget. It is a causal edge, and it is fixed by the shape of the run, not by the governor's performance target.

Why batched logging misses it.

The usual shortcut is to record the trajectory to a store after the run, or in an asynchronous batch beside it. This optimizes throughput. The agent does not wait on the write. The record arrives eventually.

Eventually is after the deadline. By the time the batch flushes, later actions have already read the earlier effects and committed their own. The examiner pulls one decision and asks what order things happened in. A batched log can show timestamps, but it cannot show that the record of the first action existed before the second relied on it. The proof of ordering is the exact thing the shortcut gave away.

Position is what makes the deadline meetable.

A record can only beat the causal edge if it is sealed on the near side of it, in the same synchronous step that authorizes the action. This is why the governor sits in the decision path rather than beside it. A control beside the path can be fast and still late. Fast is a throughput property. Late is an ordering property. They are not the same axis.

Sitting in the path is the cost that buys the guarantee. The action waits for its own record to seal before the effect is allowed to become a dependency. The wait is not overhead to be pushed toward zero. It is the mechanism by which the order is made provable.

The SLA, restated.

Governance for adaptive AI is being described this year as an engineering problem rather than a policy document. The controls have to be built into the workflow. SR 26-2 took effect on April 17, 2026 and placed agentic AI outside its scope, which hands each institution the job of defining the record its own systems keep. A record defined without a deadline is a record that can arrive after it was needed.

State the obligation as a service level. The measure is not the governor's signing speed. The measure is whether every action's record is sealed before any later action is permitted to depend on it. Time to receipt, read against the causal edge the run draws, is the governance SLA. Throughput is a separate number that matters for a separate reason.

What we are building.

Wayfinder Systems Group sits in the decision path and seals the record of each agent action in the same synchronous step that authorizes it, before the effect is allowed to become a dependency for the next action. The seal is fixed on a tamper-evident chain at the moment the action runs, so the order of the run is provable and not merely timestamped. Patents held in The Wayfinder Trust. We call her Velma.