Jev playground: try Choice, Score and Noul online

What to look at

The interesting part is not the winning option — it is the distribution underneath it. Edit an option's description and watch the probabilities move: those descriptions are the entire prompt surface of a Choice question, and writing them well is most of the skill in using Jev.

Try making the state genuinely ambiguous. A confident answer on an easy input tells you nothing; a 0.4 confidence on a borderline one tells you exactly where your threshold needs to sit.

What the Jev playground sends

Run sends two things: the state, which is the text you are asking about, and one question built from the form. The server checks both before anything is spent, then calls Jev with a 30-second timeout. For the rubric preset, what reaches the model is this, plus a model id that depends on the route:

{
  "state": "honestly whoever designed this checkout ...",
  "questions": {
    "severity": {
      "type": "score",
      "instructions": "How harmful is this content?",
      "criteria": ["Benign", "Questionable", "Harmful", "Severe"]
    }
  }
}

That is the shape TypeSafe's HTTP API takes, so a question that behaves well here carries over to your own code. The checks are about cost: the state can run to 4,000 characters and the instructions to 200. A Choice needs two to eight options with descriptions up to 200 characters each, a Score two to eight levels, and a Noul both a true and a false description.

Three presets, none of them easy

• Route a support ticket (Choice) — a customer charged for an upgrade that has not arrived, whose API calls are now failing. It is billing and technical at once. The obvious version, a double charge, comes back 1.00 to billing and makes Jev look like a keyword matcher.
• Grade on a rubric (Score) — a hostile complaint about a checkout, placed on four levels from Benign to Severe.
• Ask a yes/no question (Noul) — a password-expiry message with a link, and whether it creates artificial time pressure.

Each is a little ambiguous on purpose. A preset that came back at full confidence every time would teach nothing about what the probabilities are for.

Reading a Jev playground answer

• The headline is the chosen option for a Choice, the probability for a Noul, and for a Score the nearest level with the raw score beside it.
• Confidence appears for Choice and Score only. A Noul carries none, so the card does not invent one.
• The distribution is sorted, and a Score's probabilities are relabelled from indexes to the levels you wrote.
• The footer has the call's latency and input tokens, and a cost computed from those tokens at the published $0.042 per million; output adds nothing.

A Score of 2.68 on a four-level rubric is not a typo. It is the probability-weighted mean of the level indexes, 0 to 3, so it sits between the third and fourth levels and leans to the fourth — which is why the card shows the fraction as well as the band.

Same question, a chat model

Once Jev has answered, the buttons under the card send the same state and question to Gemini 2.5 Flash Lite or Mistral Small 3.2 24B through OpenRouter, with a plain prompt that lists your options and the JSON shape to return. The reply is then checked for type: an option that is not yours, a score out of range, or a boolean where a probability was asked for is flagged.

Our own benchmark found that sending response_format: json_schema removes those type errors entirely, so a flag here shows the weakness of a plain prompt rather than of the model. Only cheap models are offered, because the comparison runs on this site's key.

How do you compare Jev and the OpenAI Decisions API side by side?

Choose GPT-6 Luna Decisions under “Compare with” and press Run: Jev and OpenAI's model get the same state and question at the same moment, and both answers appear together with a one-line verdict on whether they agree and how sure each was. The link jev-agent.com/playground?compare=gpt-6-luna opens it ready to run. Any two of the thirteen decision models pair the same way, and a side-by-side run counts as two runs.

On the support-ticket preset, run on 2026-10-09, Jev chose technical at 27% confidence and GPT-6 Luna chose billing at 88%: a ticket that mixes a charge with an API error, read two ways. The same split across hundreds of labelled items is measured on Decisions API vs Jev.

Taking a question out of the Jev playground

When a question behaves the way you want, the instructions, the option descriptions and the shape of the state all carry over. Send them to this site's endpoint with your own key, or to TypeSafe's once a waitlist key arrives; only the host and the token change.

The playground asks one question per run, but a request can carry several, and that is where the savings are: the state travels once, however many questions come with it. Measured on the server's own clock, one question took 70 ms and 5 in the same request took 74 ms. Two things do not carry over. A threshold that looks right on three examples is not a threshold, so measure it on your own labelled data. And an alias such as jev-latest moves when a release ships, so pin a versioned id such as jev-1.13.0 once you have tuned against it.

Jev playground FAQ