A System One model writes no text: it takes a state and questions with predefined answers and returns, for each, an option plus a calibrated probability — a decision code can use directly, in a single pass.

“Think of Jev as a frontier-intelligence function call: unstructured state in, typed probabilistic decisions out.”TypeSafe AI (blogul companiei) · Introducing System One Models & Jev · 2026 · Introducing System One Models & Jev (typesafe.ai, 15 septembrie 2026) — paragraful care prezintă modelul

A System One model returns an option from your list and how sure it is; it cannot invent options, but it can pick wrong.

The name comes from Kahneman's psychology: "System 1" is fast judgement, "System 2" is slow reasoning. TypeSafe AI launched Jev (named after Jevons' paradox) on 15 September 2026 on exactly that split: an LLM reasons and generates, Jev only decides. The contract has three primitives. "Choice" picks one option from a list you supply and returns a probability for each; "Score" places the state on an ordinal scale described in words; "Noul" is a yes/no with P(true). All are evaluated in parallel, in one pass, not token by token. The training, which the company calls RLCD — Reinforcement Learning for Calibrated Decisions — optimises the probabilities to match reality: if the model says 0.9 on a hundred inputs, about ninety should be true. That is where the usefulness in software comes from: a threshold on a probability becomes an "if" that means something. The numbers, with their labels. The company claims "193.6x faster, 444.6x cheaper" than GPT-6 Astra and Fable 5.1 on four internal evaluations; stated latency is 70–500 ms, the price $0.042 per million input tokens. Independently, Every measured in September 2026 a median of 0.35 s per passage against 8.83 s for Fable, and on planted defects Jev caught six of seven, the large model all seven. "Zero hallucination" only means the schema is guaranteed: the model cannot invent an option, but it can pick the wrong one from your list.

The contract as the network sees it. A request has three parts: `state` (the text or JSON being judged), `model` and `questions`, a dictionary where each key is one question of one of the three types. The example from the TypeSafe documentation (2026), reduced to its essence:

{
  "state": "I've been trying to connect my Stripe account for 3 days and the integration keeps failing. I'm losing sales. Please help ASAP.",
  "model": "jev-latest",
  "questions": {
    "department": {
      "type": "choice",
      "instructions": "Which team should handle this",
      "criteria": {
        "billing": "Payment or subscription issues",
        "technical": "Bugs or integration problems",
        "sales": "Pricing or account questions"
      }
    },
    "frustration": {
      "type": "score",
      "instructions": "How frustrated the customer appears",
      "criteria": ["Calm, just stating facts", "Frustrated but civil", "Very angry, strong language"]
    },
    "is_urgent": {
      "type": "noul",
      "instructions": "The message conveys urgency or time-sensitivity"
    }
  }
}

The answer has the same shape as the questions, plus numbers. For "choice" you get the chosen option, the probability of every option and a confidence (the distance from the decision threshold); for "score" a value on the scale; for "noul" P(true) directly:

{
  "answers": {
    "department": {
      "choice": "technical",
      "probabilities": {"billing": 0.159, "technical": 0.840, "sales": 0.001},
      "confidence": 0.596
    },
    "frustration": {"score": 1.035, "confidence": 0.842},
    "is_urgent": {"noul": 0.999}
  },
  "usage": {"input_tokens": 312, "output_tokens": 48}
}

Two things to read in the example. "technical" at 0.84 is not "technical": it is 0.84 — if a wrong route is expensive, your threshold sits higher and the rest goes to a human. And "score" returns 1.035, i.e. between "frustrated but civil" and "very angry"; the scale is ordinal, not a label.

Where it makes sense and where it does not. It makes sense next to an LLM, not instead of one: routing and triage, classification, scoring against a rubric, checking a citation, judging a tool call before it runs, screening inputs for injections. It makes no sense for anything that needs text, arithmetic or date comparison: the TypeSafe documentation lists explicitly that the model "reads dates as text, not ordered values", that it "recognizes the shape of an answer rather than tallying", and that "injected instructions can move the answer". Context limits (September 2026): 64k tokens for state plus questions, at most 255 options per "choice".

Why it matters Whoever confuses "cannot leave the schema" with "cannot be wrong" sets a threshold on a probability they never checked.