You put a confirmation on every write. By Thursday afternoon your colleague has answered forty of them since lunch, y after y, each one a diff she stopped reading somewhere around the ninth. The forty-first drops the staging database, and it gets the same key in the same fifth of a second as the forty before it. Nothing failed. The prompt fired exactly as designed and a human was in the loop the whole time. The review that mattered received precisely the attention the thirty-nine that did not mattered received, which was none.

An alarm is worth what it stays quiet for

Picture a street where every car alarm is set to its most sensitive notch. A lorry goes past and six of them go off together. By the second week nobody at any window looks up any more, and the neighbour who switched his off entirely and the neighbour who tuned his to catch everything have arrived at the same position by opposite roads: an unguarded car, one of them silently and one of them at ninety decibels. What gives an alarm any meaning is not the noise it makes. It is the count of occasions on which it did not make any.

The alarms that keep working get there with a chain of cheap sensors in front of the expensive one — a threshold that shrugs off vibration, a tilt switch that closes only when the body of the car actually lifts. Routing attention is that chain. Blunt economic rules kill the bulk of the noise first, a classifier weighs how far an event resembles things that turned out to matter before, and a model-sized judge is spent only on the ambiguous middle. Running a language model over every event is the sensitive-notch mistake with an invoice attached.

And the verdict is not yes or no. An event can interrupt a person, be handed to another agent, or go quietly into memory. A heartbeat saying all clear is stored and never mentioned. A failing CI run is delegated, because there is something to do about it that nobody needs to watch.

Confidence alone is a poor stopping rule

The other half of the same question is how hard to stop, and it is a setting rather than a stance. A run can go full auto, stop at named gates, checkpoint between phases, step through under co-pilot, or take a policy you wrote yourself — and the useful systems change setting partway through. A research pipeline runs unattended across the phases it has done a hundred times, halts for a person to approve the experiment design, then drops into co-pilot when its confidence in the hypothesis falls under a threshold. A cost ceiling stops it outright.

The tempting shortcut is to drive all of this from the model’s own confidence and call it done. Confidence is one input and a weak one on its own, because a model can be perfectly confident about a deletion. What has to sit beside it is the risk carried by the action, how much of the run’s budget is left to spend on being wrong, and whether the step can be taken back at all. Confident and irreversible is the combination that deserves a person; uncertain and reversible mostly deserves a retry.

The same event, judged by the hour

Worthiness is not a property of the event by itself. It depends on the scene it lands in — who is free, what they can actually do about it right now, whether it is still true in an hour. An urgent request that arrives while its reader is in a meeting is not dropped and not fired through the wall either; it accumulates into a digest that gets read once. And when the routing gets it wrong, the correction is worth keeping. A person who escalates something the policy filed away has just labelled an example.

What travels also has to be worth the trip. A dispatch that amounts to have a look at this has moved the work sideways without reducing it, so an interrupt or a delegation should arrive carrying a plan and a way to check the plan was followed. That is the difference between waking somebody and reporting to them. The node that receives all this is a built part of the workflow, not a message in a channel; what this decides is which events are allowed to reach it.

Ask a person about everything and you have asked them about nothing. Choosing is the work.