agi.jetzt
DEEN

Decision Models

Jev: TypeSafe AI’s first System One Model

AvailableProduct model

In short

Jev is an AI model by TypeSafe AI, introduced on 15 September 2026. It generates no text. You give it a state and questions with fixed possible answers; for each question it returns a choice, a score or a probability of yes, each with a probability distribution. TypeSafe calls this new class of models "System One Models". The term is TypeSafe’s own, not an established industry category.

Key facts

Provider
TypeSafe AI, San Francisco
Introduced
15 September 2026
Current version
Jev 1.13 (model ID jev-1.13.0)
Access
Hosted API; also through OpenRouter and the Vercel AI Gateway
Price
0.042 US dollars per million input tokens; output tokens free
Input
Text only: a string, a JSON object or an array
Output
Choice, Score or Noul, each with probabilities
Context
64,000 tokens per request; 32,000 for the state plus the longest question

Sources: TypeSafe AI press release, 15 September 2026 (Business Wire via Yahoo Finance), TypeSafe docs: Models, OpenRouter docs: Jev · checked 10 Oct 2026

What is Jev?

Jev is a model for decisions inside software. It reads natural-language text like a language model, but it does not answer in sentences. It answers questions whose possible answers the developer has defined beforehand.

TypeSafe’s documentation puts it this way: "Like an LLM, a System One model understands natural-language input. It returns typed decisions and probabilities rather than generated text." The name Jev refers to the economist William Stanley Jevons; "System One" comes from Daniel Kahneman’s distinction between fast, intuitive and slow, deliberate thinking.

Sources: TypeSafe docs: System One, TypeSafe: Introducing System One Models & Jev

How does Jev work?

A request consists of two parts. The state is the material to be judged: a message, a record, the current state of an application. The questions say what is to be judged, each with a type and its possible answers.

Every question sees the same state and is answered independently of the others. The answers come back under the names the developer chose.

An example

Request
POST https://api.typesafe.ai/v1/systemone

{
  "state": "Help! My payouts have been failing for 3 days.",
  "model": "jev-latest",
  "questions": {
    "is_urgent": {
      "type": "noul",
      "instructions": "Does this convey urgency?"
    },
    "department": {
      "type": "choice",
      "instructions": "Which team should handle this?",
      "criteria": {
        "billing": "Payments, invoicing, refunds",
        "technical": "Bugs, outages, integrations",
        "sales": "Pricing, upgrades, new accounts"
      }
    },
    "frustration": {
      "type": "score",
      "instructions": "How frustrated is the customer?",
      "criteria": ["Calm", "Frustrated", "Very angry"]
    }
  }
}
Answer (excerpt)
{
  "model": "jev-1.13.0",
  "answers": {
    "is_urgent": { "type": "noul", "noul": 0.95 }
  }
}

Request and answer follow the examples of the TypeSafe documentation, combined into one request. The answer shows one of the three questions.

Choice, Score and Noul

Jev knows exactly three question types. Ranking or extracting values is built from them in code.

Sources: TypeSafe docs: Primitives, TypeSafe docs: API reference

TypeQuestionReturns
ChoiceWhich of these options?The option with the highest probability, a probability for every option (together 1) and a confidence value. Up to 255 options.
ScoreWhich level on this scale?A value on an ordered scale of 2 to 10 described levels: the probability-weighted mean of the level numbers, plus the distribution over the levels.
NoulIs this true?One number between 0 and 1: the probability that the answer is yes. No separate confidence value.

Probability and confidence

The confidence value is not a second opinion of the model. According to the documentation it is "a statistic computed from the probability distribution the answer already gives you": 1 when all probability lies on one outcome, 0 when it is spread evenly.

Where to set a threshold is left to the developer; TypeSafe writes that suitable values depend on the domain. And calibration, the documentation says, "is measured across groups of predictions; it does not guarantee that an individual answer is correct."

Sources: TypeSafe docs: Confidence, TypeSafe docs: System One

How does Jev differ from a general-purpose language model?

General-purpose language model
  1. Context
  2. Generation
  3. Text, code or media
Jev
  1. State + typed question
  2. Jev
  3. Choice, score or probability
  4. Software action
  1. What comes backA language model returns generated text. Jev returns one of the answers defined beforehand, with probabilities.
  2. The interfaceA language model is built for conversation with people. Jev is built to be called by programs: typed questions in, typed values out.
  3. Checking the outputText has to be parsed and validated before software can use it. A typed answer always has the expected form — which says nothing about whether it is right.
  4. Cost and timeTypeSafe charges input tokens only and states response times well under a second. Both are the vendor’s figures.

Not: probabilistic versus not probabilisticThe difference is not "probabilistic versus not probabilistic". Language models compute probabilities too, for every next token. What differs is what the model returns and which interface it is optimized for.

Where can Jev be useful?

TypeSafe names these uses itself. They describe what the model is intended for, not a measured result.

  • Classifying, routing and scoring inside existing software, where hand-written rules are too brittle
  • Checking the output of language models: scoring, guardrails, detecting jailbreaks
  • Decisions in real-time applications where a language model would be too slow
  • Many small judgments over large amounts of data

TypeSafe states just as clearly what such a model does not do: it writes no replies, produces no code and generates no explanations of its reasoning.

Sources: TypeSafe: Introducing System One Models & Jev, TypeSafe docs: System One

What are the limits?

Named by TypeSafe itself

  • It answers the question as written, not as meant.
  • It is not a calculator and does not count reliably.
  • It reads dates as text, not as ordered quantities.
  • Instructions with double negatives or several steps of indirection are answered less reliably.
  • Accuracy falls when the state holds much that is unrelated to the decision.
  • Text in the state that was written to steer the model can shift the answer.
  • With Choice, the order of the options can affect the answer; the model leans toward the first.
  • It is not trained to generate text.
  • Text only; English is the primary training language, other languages are handled less well.

Sources: TypeSafe docs: Jev 1.13 jaggedness, TypeSafe docs: Models

Found by third parties

  • The names of the options influence the decision, even when their definitions say the opposite.
  • Without an option to abstain, stated probabilities on unanswerable questions were far off.
  • Identical requests did not always return identical answers. TypeSafe documents no guarantee of determinism.

Sources: Sun et al.: Type-Safe Is Not Error-Free (arXiv:2609.26758, preprint), jujumilk3/jev-calibration-audit (GitHub)

Not published

Architecture, number of parameters and training data are not published. There is no paper, and we found no released weights. TypeSafe publishes no calibration figure and no guarantee that identical requests return identical answers.

What is documented, what is claimed, what was verified

Three kinds of statement, kept apart. The independent studies are preprints and audits from the first weeks after the launch; none is peer-reviewed, and their results point in different directions.

Documented

Checkable in documentation, a press release or a third party’s documentation.

Vendor claim

What TypeSafe says about performance, architecture and training. Not independently verified.

Independent

Measured or reported by third parties.

Why does Jev matter?

Since Jev was introduced, other providers have published models or interfaces of this kind: Cloudflare open decision models on 1 October 2026, Microsoft its Microsoft-Decision-1 on 9 October 2026; OpenAI offers a Decisions API in public beta.

At this price and speed, a judgment that used to need a language model can stand inside an ordinary code path — as often as a program needs it. Whether the judgments are good enough for a given task is what the independent studies are beginning to measure.

Sources: Cloudflare: Introducing Clef (1 October 2026), OpenAI API docs: Decisions guide, Microsoft Command Line: Microsoft-Decision-1 (9 October 2026)

Why may this matter for AGI research and agent architectures?

Agent systems consist of many small decisions around a few large ones: which tool, which model, continue or stop, good enough or hand over. A fast model with stated uncertainty fits these places — as a router, as a gate before a risky step, as a scorer of intermediate results.

It also gives research a clean object of study: when a model returns probabilities instead of text, one can measure whether they hold.

Relevance to AGI does not mean demonstrated progress toward AGI. Jev is not one of the frontier models, does not enter AGI Progress, and solves a narrower task than a generalist — it is not closer to AGI than a generalist.

Sources

System One & Decision Models: the class and seven further models →