The alert that nobody reads

The predictable failure

A Large Universe Model watching enough streams will find something to say constantly. Beliefs move all day. If every movement is reported, the reports become noise, people stop reading them, and the system's real finding — the one it existed for — arrives in a channel nobody opens.

This is not hypothetical. It is the observed history of every monitoring and alerting system ever deployed.

Why thresholds fail

Threshold alerting fires when a value crosses a line, which is wrong in both directions. It fires during expected excursions the model already anticipated, and stays silent during genuinely anomalous behaviour that happens to remain inside the line.

The better trigger

Revision magnitude relative to prior confidence. A belief the model was highly confident about, moving a long way, is surprising by construction. A belief that was already uncertain drifting further is not.

This is the natural trigger for a system that holds beliefs, and it is unavailable to systems that do not — which is much of the practical argument for the third generation over the first two.

Additional filters worth applying

Actionability. If nothing can be done before the next scheduled review, it is not an interruption.

Novelty. A belief that has been broken for a week and is already being worked is not news again today.

Aggregation. Forty estimates moving because of one congestion event is one finding, not forty.

Recipient. The same revision may be urgent for one person and irrelevant to another. Routing is part of the problem, not a separate concern.

The part with no clean answer

How much interruption a given person will tolerate before disengaging is not knowable in advance, varies by role and by week, and cannot be derived from the model.

The only real approach is feedback: track which alerts were acted on, which were dismissed, and which turned out to have mattered, then treat that as evidence about the threshold. A Large Universe Model applying its own machinery to the question of when to speak.

The conservative default

Start far quieter than seems right. A system that surfaces three things a week and is right about all three earns the standing to say more. One that surfaces thirty and is right about three never gets read again.