The Jev playground

The playground below is the real thing: it posts to the same decide endpoint you get with a key, so a decision you run here in the browser — no account, no card — is exactly what you'd ship. Type a situation, choose a primitive, and read back a typed answer with calibrated confidence.

Jev is TypeSafe AI's first System One model, launched in September 2026 by Diogo Almeida — the OpenAI researcher who co-invented RLHF and InstructGPT. Unlike a chat model, it doesn't write prose. It emits a typed structured value in a single non-autoregressive pass: one of the options you defined, a calibrated probability for each, and a confidence score — in roughly 70–500ms. The playground lets you feel that difference before you write a line of code.

What you can do in the playground

• Run a real decision with no sign-up — the browser demo calls the live endpoint, not a canned response
• Try each primitive: choice (pick one), score (rate on a scale), noul (a calibrated 0–1 yes/no)
• Start from a preset situation or type your own state in plain language
• Watch the probability bars and confidence update, so you can see how calibrated the answer is
• Compare Jev side by side with other decision models in the same panel

It's the same call you'll ship

There's no separate 'demo mode' to unlearn. The playground sends the same request shape your own code will send to the keyed endpoint — a state plus a map of typed questions — and gets back the same unified answer object. When you're ready, generate a jv_live_ key and point your code at /api/v1/decide with an identical body:

// The same request the playground makes — just add your key
const res = await fetch("https://jevtypesafeai.com/api/v1/decide", {
  method: "POST",
  headers: {
    "Content-Type": "application/json",
    Authorization: `Bearer ${process.env.JEV_KEY}`, // jv_live_...
  },
  body: JSON.stringify({
    state: "A user wrote: 'this is the third time it broke, I want a refund.'",
    questions: {
      sentiment: { type: "choice", instructions: "Classify the tone.",
        criteria: { angry: "frustrated/upset", neutral: "matter-of-fact", happy: "positive" } },
      urgency: { type: "score", instructions: "How urgent, 1 (can wait) to 5 (now)?" },
      refund: { type: "noul", instructions: "Is the user explicitly asking for a refund?" },
    },
  }),
});
const { answers } = await res.json();
// answers.sentiment.choice -> one of your keys
// answers.urgency.score    -> a number on your scale
// answers.refund.noul      -> a calibrated 0-1 probability

answers.sentiment.choice is locked to the keys you declared, so there's no invalid label to handle; answers.refund.noul is a calibrated probability RLCD training makes trustworthy. Every question comes back in the one call — the playground just shows you that answer object rendered as bars.

The three primitives, live

• choice — pick exactly one of a fixed set of options you define (route, classify, select)
• score — rate on an ordered scale when you want a graded threshold (priority, risk, quality)
• noul — a single calibrated 0–1 probability for a yes/no question ('is this safe?', 'did it succeed?')

From playground to production

Trying it in the browser is free and needs no account. When you move to your own key, output is free and input is billed per million tokens on a sliding scale that gets cheaper the more you prepay — in practice a single decision costs well under a cent. Current tiers are on the pricing page.

Try your own situation

The fastest way to understand Jev is to paste a real decision from your own product — a message to classify, two answers to pick between, an action to vet — and watch it come back typed. Edit the state in the playground above and run it again.

FAQ

See also: Get a Jev key · Is Jev free? · Jev examples · How does Jev work? · Pricing