Jev examples

Jev, in plain terms

Jev is a System One model from TypeSafe AI. You send it the state of your application plus typed questions, and it answers each question with a decision and a calibrated probability, in a few hundred milliseconds, without generating a single word of text.

Not a chat model

A language model writes an answer token by token, and you parse the prose afterwards. Jev skips the writing. You define every possible output up front, and it returns one of those outputs together with how sure it is. That single difference explains most of its properties.

Typical LLM callJev
OutputFree text you must parseTyped values matching your question
LatencySeconds70 to 500 ms per request
PriceInput and output tokens$0.042 per million input tokens, output free
HallucinationPossible, needs validationImpossible by construction: it can only pick from your options
UncertaintyHidden in the wordingA probability on every answer

What it gives up is prose. Jev cannot explain, summarize, or write. It decides, classifies, routes, and scores. TypeSafe calls this class of model System One, after the fast, intuitive mode of thinking.

State: what Jev looks at

Every request carries a state: the thing to evaluate. It can be a plain string, a JSON object with named fields, or an array of text values such as a chat transcript. Jev is text-only. Images, audio, and video are not supported, so a perception step has to turn them into text first.

"state": { "message": "My card was charged twice", "order_id": "A-104" }

Questions: the three primitives

Alongside the state you send one or more named questions. Each has a type, a line of instructions, and, for two of the types, criteria that define the allowed answers. Jev answers all of them in one pass.

noul: a yes or no question

Returns the probability that the answer is yes, from 0 to 1.

"is_urgent": { "type": "noul", "instructions": "Does this convey urgency?" }
→ { "type": "noul", "noul": 0.95 }

choice: pick one option

You give a map of option keys to descriptions, up to 255 of them. Jev returns the winning key, a probability for every option, and a confidence score.

"department": {
  "type": "choice",
  "instructions": "Which team should handle this?",
  "criteria": { "returns": "Exchanges, wrong or damaged items",
                "billing": "Charges, invoices, payment problems" }
}
→ { "type": "choice", "choice": "returns", "confidence": 0.98,
    "probabilities": { "returns": 0.99, "billing": 0.01 } }

score: position on an ordered scale

You give 2 to 10 level descriptions, lowest first. Jev returns a position on that scale as a decimal, the probability of each level, and a confidence score. A score of 1.43 means "mostly level 1, leaning toward 2".

"severity": {
  "type": "score",
  "instructions": "How severe is the reported issue?",
  "criteria": [ "Cosmetic", "Degraded, workaround exists", "Blocking" ]
}
→ { "type": "score", "score": 1.43, "confidence": 0.35,
    "probabilities": { "0": 0.0, "1": 0.57, "2": 0.43 } }

Confidence and calibration

Every answer comes with a probability distribution, and choice and score answers add a confidence from 0 to 1 that reflects how concentrated that distribution is. A confidence of 0.98 means one option dominates. A confidence of 0.35 means Jev is torn.

The point of calibration is that these numbers are meant to be trusted literally. When Jev says 0.7, it should be right about seven times in ten. That is what lets you write code such as "auto-approve above 0.9, send to a human below 0.6". TypeSafe describes the training as reinforcement learning aimed at calibrated decisions rather than at fluent text.

Where it fits

How this site calls it

The examples here go through OpenRouter's Decisions API, which hosts Jev under the id typesafe/jev-1.13. The API key stays on this site's server. Every run shows a panel with the time and cost of each step, so you can see what a decision actually costs: a typical request is a few hundred tokens, which is a few thousandths of a cent.

Jev was released on 15 September 2026. An open-weights alternative called Laya, from Convai Innovations, answers the same three question types and can be self-hosted.

Try it