Decisions as Events

This is a short note, and the idea in it isn’t finished. Most of my attention has gone to the observational side of this, the data and what it means. This is about the other side, and I know it less well.

In the last essay I described an ambient agent: you walk away, it stays on, and it tells you when something looks wrong. An anomaly alert is the easy example. A metric moves, and the agent has to decide whether that’s worth your attention.

It makes that decision constantly, and nearly always the answer is no. That’s an odd job to give a large language model. Asking one every few minutes is slow and costly, and what comes back is a paragraph. I also warned last time that an agent that pings people too often gets muted. So a lot rides on the quality of this one small decision.

A new kind of model looks built for it. TypeSafe’s Jev, in early access since September, and Microsoft’s Decision-1, which followed in October, are called decision models.¹ You give one some state and a question with a fixed set of answers. It returns a typed answer with a probability for each option, and no prose. Both are priced at a few cents per million tokens. TypeSafe calls the category System One, after Kahneman: the fast reflex, as against slow deliberation.

Where I think it sits

Here is my best guess at the embedding, in three layers.

Measuring stays in SQL and statistics. Jev, at least, isn’t meant for exact arithmetic or comparing against thresholds. Whether a number is three standard deviations out is a query.

Judging is the decision model. Given the anomaly, the metric’s definition, and what else is known (a launch on Tuesday, a holiday), does this matter? Who owns it? Is the proposed fix safe to apply without sign-off? Each is a question with a short list of answers.

Explaining is the large model, and only when the judgment says to. It writes the message to the owner and drafts the change.

Why the event matters

Each judgment is a row: the question, the options, the probabilities, the threshold applied, the model version.² It can sit in the warehouse beside the data it was about.

That closes a loop I left open. When an owner acts on an alert or dismisses it, that’s a label. The guidance for Jev is to set thresholds from your own labeled data, in three bands: act, review, escalate. So the people being interrupted are the ones tuning how often they’re interrupted, and they do it by responding. That’s the same shape as the self-healing semantic layer, applied to attention.

What I don’t know

Calibration is the vendors’ claim, and the benchmarks so far are mostly their own. Amplitude tested Jev and found larger LLMs more accurate on its benchmark. I don’t know how much context a judgment like “does this matter” needs, or whether a model this small can use it. And I haven’t built any of this. I’m writing it down because it looks like the missing piece, and I want to find out whether it is.

Footnotes

1. Kenny Rogers, “What Is Jev? TypeSafe’s Decision Model Explained for Developers,” OpenRouter, and Tom Smith, “Microsoft Takes Aim at Jev With a Fast, Cheap Decision Model,” Techstrong.ai, October 13, 2026. Both list a price of $0.042 per million input tokens. The Techstrong piece is the source for the Amplitude result. ↩

2. From the OpenRouter explainer, which suggests logging the request id, the question names, the probabilities, and the threshold your code applied, and keeping the state itself out of the log because it may hold customer data. ↩