What Is Jev?

A decision model, not a text model

Most AI APIs return generated text that your code then has to parse. Jev skips the text. A request carries a state and a map of typed questions; the response carries one typed answer per question — a yes/no probability, a pick from your options, or a position on your scale. TypeSafe calls Jev the first System One model, after Daniel Kahneman's name for fast, intuitive judgment.

Jev was trained with RLCD, Reinforcement Learning for Calibrated Decisions, so a higher probability is meant to track higher accuracy. It cannot write text at all: the answer object is the whole output.

Who makes Jev

Jev is built by TypeSafe AI, a San Francisco company founded in 2024 by Diogo Almeida, Erik Gafni, and Sasha Sheng. Almeida, the CEO, spent about four years at OpenAI working on RLHF, InstructGPT, ChatGPT, and GPT-4 before leaving in 2024. Jev is not an OpenAI model.

Jev was released in limited early access on 15 September 2026, alongside a US$40 million seed round led by DCVC. Its weights, parameter count, and architecture are not published. Jev is not fine-tuned per customer — the same weights serve every account — and TypeSafe says the model is not trained on customer requests or responses.

System One vs a language model

A language model completes text; Jev scores decisions. There are no paragraphs to read and no tokens to stream — the typed answer object is the output. TypeSafe quotes 70 ms to 500 ms of end-to-end latency, which makes Jev practical inside synchronous paths such as request routing, moderation gates, and form validation.

Input is text only: a string, a JSON object, or an array of text. Jev cannot see images, audio, or video. English is the primary language and the most accurate; other languages, including Chinese, Japanese, and Korean, work but not equally well.

The three question types

Every question sets a type and instructions, which may be a string, an object, or an array. The id you choose is only a label — the answer comes back under the same key.

• Noul is a yes/no question. The answer is noul, a probability between 0 and 1 that the statement is true. An optional criteria object describes what true and false mean.
• Choice picks one option from a map you define, up to 255 options on the official API. The answer includes choice, a probabilities entry for every option, and confidence.
• Score places the state on an ordered scale you write, 2 to 10 levels, lowest first. The answer includes a fractional score that can land between levels, a legend, probabilities, and confidence.

{
  "state": "Help! My payouts have been failing for 3 days.",
  "model": "jev-latest",
  "questions": {
    "is_urgent": { "type": "noul", "instructions": "Does this convey urgency?" },
    "team": {
      "type": "choice",
      "instructions": "Which team should handle this?",
      "criteria": { "billing": "Payments and refunds", "technical": "Bugs and outages" }
    },
    "frustration": {
      "type": "score",
      "instructions": "How frustrated is the customer?",
      "criteria": ["Calm", "Frustrated", "Very angry"]
    }
  }
}

What Jev is bad at

TypeSafe publishes a jaggedness note for Jev 1.13, last reviewed on 2 October 2026. The documented weak spots:

• Literal reading — Jev tends to take text at face value.
• Math and counting — keep arithmetic in your own code.
• Date and time comparison — compute relative dates yourself.
• Indirection — chains of references inside the state can confuse it.
• Large states full of irrelevant detail — filter before you send.
• Adversarial content — hostile text inside the state can skew answers.
• Contradictory instructions or criteria produce unreliable results.
• Choice option order — Jev leans toward the first option, so reorder options to check consistency.
• Generation — Jev cannot write text; use a generative model for that.

The practical advice matches the list: do arithmetic and date math in code, strip irrelevant fields from the state, and shuffle Choice options when a ranking matters.

How to try Jev

On this site, sign in and create a key — the playground at /dashboard/home can issue one for a signed-in account and run a decision in the browser. A successful call uses 1 credit, so add a small one-time top-up on the pricing page first. The official TypeSafe SDKs work here too: point the base URL at https://jev-api.org and send the standard System One body to /v1/systemone.

TypeSafe also runs its own console with per-token billing at console.typesafe.ai

FAQ