Explainer

One Look

A decision model does not write. You give it a text and a list of questions, and it answers each with a probability. What that is, why there are suddenly so many, and what five of them did with fifty messages.

A customer writes to a bike shop. Two weeks cycling on the coast, lovely weather, and one question: which panniers should she buy?

Halfway down she mentions that her husband’s handlebar came loose on the last descent, and he went over the bars. Three stitches. Then back to the panniers.

Somebody has to catch that sentence today, in an inbox of hundreds. You could have a chatbot read every message, and it would write you a thoughtful paragraph about each one. What you wanted was one word: yes.

This post is about the kind of model that gives you that one word. It explains what a decision model is, where the name System One comes from, why so many appeared in the last month, and what happened when we gave fifty messages to five of them.

What a decision model is

A chat model writes text. A decision model does not. You hand it some content and a list of questions, and for every question it returns a probability.

Every question has one of three fixed shapes.

The vendors use different words for the same three things. A yes/no question is a noul in TypeSafe’s System One format and a predicate in OpenAI’s Decisions API. The idea is the same.

Here is one real call, in outline and trimmed to two of its six questions. The content goes in as state, the questions as a named list.

{
  "state": "From: Dorian Ashcombe\nSubject: Brake failure\n\nMy brakes failed on the way to work and I broke my arm. I am contacting a solicitor about this.",
  "questions": {
    "safety_incident": {
      "type": "noul",
      "instructions": "Does the message report a product failure that injured someone or could have injured someone?"
    },
    "urgency": {
      "type": "choice",
      "instructions": "How soon does this message need a reply?",
      "criteria": {
        "routine": "No deadline is mentioned and nothing is blocked",
        "this_week": "A deadline within days is mentioned",
        "today": "Someone is blocked or at risk now, or today or tomorrow is named as the deadline"
      }
    }
  }
}

And this is what Jev 1.13 answered, 272 milliseconds later:

{
  "safety_incident": { "type": "noul", "noul": 0.97 },
  "urgency": {
    "type": "choice",
    "choice": "routine",
    "confidence": 0.49,
    "probabilities": { "routine": 0.66, "this_week": 0.05, "today": 0.28 }
  }
}

There is no text to parse and no essay to read. Software can use the answer as it is: if safety_incident > 0.6, page the workshop.

Look at the second answer, though. A broken arm and a solicitor, and the model says routine. Keep that in mind. We come back to it, and it was our fault.

Where the name System One comes from

The name is borrowed from psychology. In Thinking, Fast and Slow, Daniel Kahneman describes two ways of thinking.

• System 1 is fast and automatic. You see a face and you know it is angry. You did not work that out.
• System 2 is slow and deliberate. Seventeen times twenty-four. You can do it, but you have to go through the steps.

A chatbot answers by writing, one word after another, each built on the last. That loop is what lets it work through steps, the way System 2 does. It is also why every answer costs time and money: each word is one more turn of the loop.

A decision model takes one look. It reads the content once and answers all your questions at the same moment, as numbers. TypeSafe, the startup that coined the term System One model, describes its model as putting out all probabilities in parallel instead of generating them token by token.

The rule of thumb: count the steps

When do you need which? The clearest rule we found comes from a short video by the channel Prompt Engineering: count the steps.

• Is this a refund request? The answer is visible in the text. One look. A decision model will do.
• Did we promise this customer a refund last month, and was it ever paid? That needs things you have to go and find first, and then compare. That is a job for a model that writes, or for an agent.

Why there are suddenly so many

The timeline is short.

• September 2026. TypeSafe ships Jev and calls the category System One.
• 29 September. OpenAI names a Decisions API at its developer day, as a limited preview.
• 6 October. The Decisions API opens to all developers.
• 11 October. OpenRouter alone lists 19 decision models from 12 vendors.

The reason is the price of a glance. These are our own measurements from one laptop on 11 October, six questions per call, twelve messages per model. It is a small sample, so read it as an order of magnitude.

At that price you can check everything: every email, every ticket, and every step an AI agent is about to take. That last use is the one the vendors talk about most. TechCrunch describes a demo that checks each action of an agent against its task, and blocks the ones the model is confident are wrong.

What is not new

It is worth being dry about this. Reading a text once and sorting it into a category is not new. Trained transformer classifiers have done it since 2018, and zero-shot classifiers, which take the labels as plain text, since 2019.

If your labels never change and you have examples to train on, a small trained model can still be cheaper, and sometimes better. The same Prompt Engineering channel ran that comparison and reports a 22-million-parameter local model scoring 93% on a banking benchmark where Jev scored 80%. That is their test, not ours.

What is new is the question. You write it in plain language, with a description for each answer, and you can change it this afternoon. There is nothing to train.

We tried it: fifty messages, five models

We wrote one week of a support inbox for a bike shop that does not exist: 36 everyday messages in English, German and French, and 14 messages written to be difficult. Then we asked six questions of every message.

One person wrote an answer key before any model ran, and did not change it afterwards.

The everyday inbox. Jev 1.13 agreed with the key on 201 of 204 answers. The three differences are worth reading. It called a reporter’s question about brake failures a safety incident (69% yes). It rated a price-match question as needing a reply today, at 46% confidence. And for a fourth email threatening the small claims court, it answered annoyed at 47%, just under our 50% line, so nothing was recorded.

The hard cases. The five models gave the same answer on 70 of 84 questions.

Three things in those results are more useful than the scores.

1. A threshold does not travel between models

Back to the letter about the panniers. Four of the five models gave the buried injury 94% or more. The fifth, Clef Flash, gave it 59%. Our line was 60. It missed by one point.

That is not an accident of one letter. On the 31 pick-one and scale answers where all five models agreed, Clef Flash reported the lowest confidence of the five on 24. Its median was 0.80 where the others’ was 0.97. The answers were the same; the numbers were not.

So a threshold tuned on one model drops correct answers on the next. Set it per model, on messages you already know the answer to.

Here is the same letter in the findings list. Four models left a safety_incident finding. The fifth left only the team it would route the message to.

2. The models follow what the options say, not what you meant

Now the broken arm. “My brakes failed on the way to work and I broke my arm. I am contacting a solicitor about this.” Three of five models rated its urgency as routine.

That message also carried an order to the model, so the order was the obvious suspect. We ran the message again without it.

The three that said routine still said routine. The cause was us. We had described the top level, today, as “someone is blocked or at risk now, or today or tomorrow is named as the deadline”. An arm that is already broken is neither blocked nor at risk now. The models read the options literally, and the options did not say what we meant.

3. Text in a message can move an answer without being obeyed

Look at the first row of that table again. No model obeyed the order. But Jev answered today with the order in the message and routine without it. The injected sentence pushed its answer in the opposite direction to the one it asked for.

Nothing obeyed the order, and it was not ignored either. If the content you judge can be written by strangers, test with and without the strange parts.

Where this sits in Classifyre

We built this test in Classifyre, where a decision model is one kind of detector. You pick a model and write the questions; under each question the form shows which findings it will record.

A scan then turns every answer that matters into a finding, with the name and the priority you gave it.

Opening a finding shows all six answers of that call with their confidence, including the ones that stayed below the line and recorded nothing. That list is where every number in this post was read from.

The decision detector docs have the setup and the full schema. The models are called through LiteLLM, so the same questions run on any of the providers above, or on a model you host.

Writing questions that work

Most of what went wrong in our test was in how we wrote the questions. These are the habits that would have saved us.

• Ask about something visible in the text. “Does the writer ask for a manager?” works better than “Is this serious?”.
• One thing per question. Split “Is this urgent and about billing?” in two.
• Describe every answer, and read the descriptions as a pedant would. That is how the model reads them.
• Add a catch-all, such as other, so that content which fits nothing is not forced into a real answer.
• Set thresholds from messages you know the answer to, and set them again when you change the model.

What a decision model will not do

• It gives no reasons. You get a number, never a because. If you need the explanation, ask a model that writes.
• It cannot work through steps. One look is all it takes, and all it has.
• Its confidence is the model’s own. The same answer came back at 0.40 from one model and 1.00 from another.
• Fifty messages are a demonstration, not a benchmark. They show how these models behave. They do not show which one is best.

So anything near the line still goes to a person.

Sources

• TypeSafe, Introducing System One Models & Jev, and its documentation.
• OpenRouter, Jev guide and its list of decision models, counted on 11 October 2026.
• TechCrunch, OpenAI’s Jev clone could help the frontier lab stop its swarming agents, 30 September 2026.
• Prompt Engineering, System 1 vs System 2 AI models: what’s actually different? and OPEN JEVs are Here!.
• Daniel Kahneman, Thinking, Fast and Slow, 2011.

More from the blog

Business blog

Oct 11, 2026

One Map: EmbeddingGemma 2 Explained, and What It Did on Four Processor Cores

EmbeddingGemma 2 puts text, pictures, sound and video into one vector space, and it is small enough to run without a graphics card. How it works, what came before it, and what we measured when we ran it on a CPU against an invented company's shared drive.

Read article→

Business blog

Oct 10, 2026

Your Laptop Remembers: How to Find Leaked Secrets in Screenshots, .env Files and Old Notes

An API key in a screenshot, a .env file in an abandoned side project, a text note full of passwords. A step-by-step walkthrough of finding every leaked secret on your own laptop with one Docker command, read-only, without anything leaving the machine.

Read article→

Business blog

Aug 29, 2026

We Deleted Our Prettiest Screen: From Fingerprints to Near-Duplicate Review

Why we replaced the fingerprints similarity graph with a duplicate review queue — what the canvas cost us, what changed for cases, and what the new numbers actually mean.

Read article→