What Jev is

Jev is TypeSafe AI’s System One model: it does not generate text; it only makes structured judgments.

A chat model is an autoregressive text generator: it writes natural language one token at a time. In theory that includes JSON.

Getting structured JSON from an LLM still means emitting text, then constraining it with prompts and decoding strategies.

Jev skips text generation. It does not need elaborate prompts or decoding constraints. Formatted output is a first-class part of the design.

Input

state can be all the context as text, or a ticket that grows in a conversation. If the problem is Markov, sometimes the current state is enough (which is why we use Jev to drive a browser or play a simple board game).

questions: Jev can handle several questions on one program state. As the number of questions grows, latency does not show a drop — that is its parallel decoding.

Output

Jev returns JSON with a fixed set of fields, including probabilities and confidence scores.

A request and response that could happen in real life

Wednesday, 9:30 p.m. You get home, open the fridge: two eggs, half a cabbage, leftover rice. It is raining. What do you do?

If you cook, fried rice is clearly more favorable than delivery or going out.

Now we ask Jev to decide.

First, define the background.

state = (
  "It is 9:30 on Wednesday night. I just got home. The fridge has two eggs, half a cabbage, and a bowl of leftover rice."
  "It is raining."
)

state gives Jev the background and prior knowledge it needs. A detailed, clear state is key to a high-confidence decision.

With that background, ask Jev the question.

{
  "dinner": {
    "type": "choice",
    "instructions": "How does dinner most likely get solved tonight?",
    "criteria": {
      "Cook": "Cook with what is already at home",
      "Order delivery": "Order on the phone and wait for it",
      "Go out": "Walk to a nearby place",
      "Skip it": "Skip this meal"
    }
  }
}

dinner is a short name for this question — settling dinner. You can write eat, hungry, or similar.

type: choice is one kind of question. Think of it as a multiple-choice item on an exam.

instructions define the question. criteria provide the closed set of options.

You build criteria by giving each candidate a text description and a unique key.

With state and questions ready, a request to the official endpoint looks like this:

curl https://api.typesafe.ai/v1/systemone \
  -H "Authorization: Bearer sk-••••••••" \
  -H "Content-Type: application/json" \
  -H "Accept: application/json" \
  -d '{
    "model": "jev-latest",
    "state": "It is 9:30 on Wednesday night. I just got home. The fridge has two eggs, half a cabbage, and a bowl of leftover rice. It is raining.",
    "questions": {
      "dinner": {
        "type": "choice",
        "instructions": "How does dinner most likely get solved tonight?",
        "criteria": {
          "Cook": "Cook with what is already at home",
          "Order delivery": "Order on the phone and wait for it",
          "Go out": "Walk to a nearby place",
          "Skip it": "Skip this meal"
        }
      }
    }
  }'

Response:

{
  "model": "jev-1.13.0",
  "answers": {
    "dinner": {
      "type": "choice",
      "choice": "Cook",
      "confidence": 0.76,
      "probabilities": {
        "Cook": 0.82,
        "Order delivery": 0.18,
        "Go out": 0.0,
        "Skip it": 0.0
      }
    }
  },
  "usage": { "input_tokens": 430, "output_tokens": 56 }
}

Try it!

Jev prices output as free, but still counts output_tokens. That suggests a decoding process is still there.