Skip to content
Jev StudioDeutschSign in

Ask a question.Read a probability.

Jev Studio is where you word the questions your code asks Jev, watch their probabilities move, prove them on labeled examples, and take calibrated decisions into production.

Sign inSee how it worksInvitation only.

What is Jev Studio?

Jev Studio is a web workbench for Jev, the decision model by TypeSafe AI. Jev reads a text and answers typed questions (yes or no, pick one, or a position on a scale) with calibrated probabilities instead of generated words. Jev Studio is where you word those questions, prove them on labeled examples, and take them into code.

What is Jev?

Jev (typesafe-ai/jev) is a System One decision model. It answers a fixed set of typed questions about one text in a single pass and returns a probability and a confidence for each answer. It writes no free text and gives no reasoning, and it is billed on input tokens only.

TypeSafe documentation

How it works: the calibration loop

Four steps take a question from a hunch to a decision your code can trust.

  1. 01

    Word the question

    Write the text Jev should read and ask it yes or no, pick one of your options, or place it on ordered levels. Everything saves as you type.

  2. 02

    Watch the number move

    Every ask shows the probability, how decided the answer is, and how far it moved since the last wording. Drag a yes-or-no threshold on the answer itself.

  3. 03

    Prove it on examples

    Add texts with the answer each should get. Jev Studio shows how many a question gets right, which it misses, and the threshold that separates them best. Compare two wordings side by side.

  4. 04

    Take it to code

    Copy the request, decision code with your thresholds, and your labeled examples as test cases. What the studio shows is what your code decides.

Watch Jev answer

Three texts, three questions each, one of every type. Pick a text or let it play.

We waited forty minutes for a table we had booked, and the risotto arrived lukewarm. That said, our waiter was genuinely lovely and comped the dessert without us asking. Not sure we will be back.

375 ms for all three questions, 454 input tokens

Recorded answers from jev-1.13.0 on 2026-09-29. This page never calls Jev.

  1. Should the owner answer this review personally within a day?

    Yes or no
    0.82Yesat threshold 0.60
  2. What is the overall sentiment of this review?

    Pick one
    0.88mixed
    • mixed 0.88
    • negative 0.12
    • positive 0.00
  3. How many stars would the author most likely give?

    Scale
    2 stars1.26 / 4

Jev or a prompted LLM as a classifier?

Both can label a text. They differ in what they hand your code.

AspectJevPrompted LLM
What comes backJev:A calibrated probability per answerPrompted LLM:Generated text you parse
Question typesJev:Yes or no, pick one, ordered scalePrompted LLM:Anything the prompt describes
Where the decision livesJev:An explicit threshold in your codePrompted LLM:Implicit in the prompt wording
BillingJev:Input tokens onlyPrompted LLM:Input and output tokens
Best forJev:Branching code on options you know in advancePrompted LLM:Writing text and open-ended reasoning

Measured, not promised

Measurements from Jev Studio and its harness, September 2026.

for one question, asked through the deployed studio (29 Sep 2026)
476 ms
to evaluate a set of five labeled reviews, four calls in parallel (29 Sep 2026)
656 ms
routing decisions matched human labels in a pilot of the owner's agent harness (20 Sep 2026)
24 of 24
per million input tokens, TypeSafe list price (20 Sep 2026)
$0.042

Questions

Who can use Jev Studio?

Jev Studio is invitation only. The admin creates a one-time link that works for seven days; the invited person signs in with GitHub or sets an email and password. Each person sees only their own studies.

Which question types does Jev answer?

Three: yes or no (noul), which returns one probability; pick one (choice), which returns a probability for each option; and scale (score), which returns a position on ordered levels.

How do I choose a threshold?

Label a few examples with the answer each should get and evaluate the set. For yes-or-no questions Jev Studio suggests the threshold that gets the most examples right and stays farthest from every one of them, so small wording changes do not flip answers.

What does asking Jev cost?

Jev is billed on input tokens only; the list price was 0.042 dollars per million input tokens in September 2026. A single question over a short review cost about a thousandth of a cent. Jev Studio adds a daily limit per person.

Where are texts and studies stored?

Studies live on the server that runs Jev Studio, on a private machine in Germany. The texts and questions of each ask are sent to TypeSafe, a US processor, to be answered.

Does Jev replace a language model?

No. Jev decides between options you know in advance; a language model writes text and explores open questions. Use Jev where your code branches, and a language model where text is the output.

Ask a question.Read a probability.

Sign inInvitation only.