Jev - A Decision Model

What does Jev return?

A typed answer and a probability, not a paragraph. You send a state (the thing being judged) and one or more questions. Jev evaluates each question against that state and returns a value your code can branch on. Choice, Score, and Box also return a confidence. A yes/no answer does not: its probability is the whole signal.

What is a System One model?

TypeSafe’s name for a model built to make fast, structured decisions for software. It still reads natural language. It does not write a reply, produce code, or explain itself. Jev is the first model in that class, and the flagship. The name points at Daniel Kahneman’s fast judgment, not at a smaller chatbot. The lore page goes into both names.

When do I use Noul, Choice, or Score?

Match the question to the shape of the answer you need. Box is a further type, for a place in an image rather than a yes, an option, or a score.

• Noul is a yes or no. The answer is the probability that the answer is yes, from 0 to 1. Use it when the useful signal is “how true is this?”
• Choice picks one option from a set you define, with no order between the options. You get the selected option, a probability for each, and a confidence.
• Score places the state on an ordered scale whose levels you describe. The score can sit between two levels. You also get a probability for each level, and a confidence.

Why does this site say Boolean?

Same question, two names. TypeSafe calls the yes/no primitive Noul. The Laravel AI SDK exposes that classification as a boolean, and this site’s PHP snippet reads it with answer('boolean')->isTrue(0.8). The JavaScript snippet reads noul. On this page, yes means the probability is above 0.8.

What does 0.85 mean?

For a yes/no question, 0.85 is the probability of yes. It is above this site’s 0.8 line, so the answer is still yes, just less sure than 0.98. Near 0 is a strong no. Near 0.5 means the model gives yes and no about equal weight. That is uncertainty, not a middle category. If you wanted “medium,” ask a Score with a middle level.

TypeSafe trains these probabilities to be calibrated across many predictions: of the answers given as 0.9, about nine in ten should be a yes. Calibration describes a set of answers. It does not promise that one particular answer is right. The threshold is yours. 0.8 is this site’s line, not a rule of the model.

What is confidence, and why is it missing on yes/no?

Confidence summarizes how peaked a Choice or Score distribution is. A Choice that puts almost all of its probability on one option is confident. A Choice split across three options is not, even if one option is slightly ahead. Use it to decide whether to act or to hand the case to a person. A Noul answer has no separate confidence field. The probability already says how strongly the model leans yes. A Box answer has its own confidence, from 0 to 1. TypeSafe has not published a threshold for it.

Can a score land between two levels?

Yes. A Score is a probability-weighted position on your scale. If the levels are 0, 1, and 2, a score of 1.4 sits between the middle level and the top one. The distribution tells you how the weight was split. This site’s frustration scale is Calm, Annoyed but polite, and Angry or threatening to leave.

One big question, or several small ones?

Several small ones. Each question should be a judgment a knowledgeable person could make in a few seconds. “Rate this pitch” hides three judgments. Ask about market, feasibility, and differentiation, then combine the scores in your code. When the weighting changes, you change a coefficient, not a prompt. Questions in one request are independent and run in parallel against the same state. Adding one barely changes the wait. If a later question truly depends on an earlier answer, that is a second request.

Does Jev explain itself?

No. There is no reasoning trace and no written justification. The answer is constrained to the options or levels you supplied, so it cannot invent a fourth team or a level you did not define. If you need prose for a person to read, that is a different model, and Jev can sit in front of it as a check.

What can I send?

Text, and images when you opt in. A string, a JSON object, or an array of text is the ordinary state. An image is listed once in the request’s assets, and a reference to that asset can sit in the state, in a question’s instructions, or in its criteria. Any request with an image sends the header TypeSafe-Danger-Zone-Images: true. Jev 1.13 and Jev 1.14 both accept images that way. Audio and video are still not accepted. When the state has several parts, name the field in the question, for example the customer’s message rather than “the ticket.” Question IDs are for your code. They are not sent to the model, so the instructions have to carry the whole question.

When is a language model the better tool?

When the product is the text: a reply, a draft, an explanation, code a person will read. Jev is the decision inside the workflow: route, score, verify, gate. TypeSafe’s published comparison puts a frontier chat model at seconds to minutes, and a System One call in a band of about 70 to 500 milliseconds, with output tokens they describe as free. Those figures are theirs, from the launch post, not a measurement of this site.

What is a Box?

A Box finds where something is in an image. The answer is one bounding box: top, bottom, left, and right. Top and bottom are fractions of the height, and left and right are fractions of the width, each from 0 to 1, measured from the top-left. It also returns a confidence from 0 to 1. There is no separate “not found” answer, so a missing target still comes back as four edges. A Box needs an image in the request, the header TypeSafe-Danger-Zone-Images: true, and the model jev-1.14.0 or the alias jev-preview. Jev 1.13 does not answer Box questions. TypeSafe calls images and Box a danger zone: opt in per request, and they do not recommend them for production yet.

How long can a request be?

Two token limits. The default is 64k for the whole request and 32k for the state plus any one question. The header TypeSafe-Danger-Zone-Long-Context: true raises both to 262k. It works on Jev 1.13 and Jev 1.14. TypeSafe says a request that uses it is slower and less accurate than one inside the default limits, and that the price per token does not change.

Which model id should I send?

Jev 1.14.0 was released on 14 October 2026. The model id is jev-1.14.0. The alias jev-preview points at that id. The alias jev-latest still points at jev-1.13.0. The models page calls jev-latest the most recent stable release. The 14 October post said jev-latest would move to jev-1.14.0 a week after the announcement. Input is $42 per Btok, or $0.042 per Mtok, on both versions. A Btok is a billion tokens and an Mtok is a million. Output tokens are free. The price did not change with this release.

How does Run work here with no API key?

The public site has no OpenRouter key, so Run replays a recorded answer and lets the numbers move a little, without changing the decision. A strong yes stays a strong yes. The unmarked question stays above 0.8 and below 0.9. A statement stays a no. Set OPENROUTER_API_KEY in a checkout of the repository and Run calls Jev.