Independent guideAn independent guide to Jev by TypeSafe AI

A practical guide for curious people

What is Jev?

Jev is an AI model that classifies text and checks it against questions you define.

For example, give it a customer message, ask “What kind of issue is this?” and define the possible categories. Jev picks one and returns a probability for each.

Independent guide

Workflow illustration

Sort customer feedback

  1. 1 · Give Jev a message

    I was charged twice for my monthly subscription. Can someone help me check the two payments?

  2. 2 · Set a question and possible answers

    What kind of feedback is this?

    You define the possible answers. This example uses the six categories below.

  3. 3 · See each answer’s probability
    BillingMost likely
    94%
    Bug or fault
    3%
    Help with usage
    1%
    Feature request
    1%
    Other
    1%
    Positive feedback
    0%

Illustrative probabilities, written by us. These are not recorded Jev results or measured accuracy.

Try this customer message

What people are exploring

Three public experiments, explained in plain language. Reported by their authors; not rerun by this website.

Vercel

Review a command before it runs

Vercel evaluated Jev as a candidate reviewer for commands in fx auto mode. A model judgment does not itself grant permission to run a command.

Author-reported experiment · Not rerun here

Explore this case
Every

Check a piece of writing

Every tested articles against explicit writing checks. The results help the writer choose which articles and checks to review; missed issues still need human attention.

Author-reported experiment · Not rerun here

Explore this case
Good Start Labs

Check an answer against a rubric

Good Start Labs reported experiments assessing game tasks and research answers. Disagreement is a reason to look closer, not proof that the verdict is correct.

Author-reported experiment · Not rerun here

Explore this case

A judgment needs a next step

Jev does not write the reply, approve a refund or guarantee that a decision is correct. People and software decide what happens next.

Evidence and limits
03

Choose how far to go

You can understand the idea without knowing an API. The technical guides are here when you need them.

A decision model for your software

Jev is an AI model from TypeSafe AI for making structured judgments inside software. Give it information and a set of well-defined questions; your application receives answers it can use to choose the next step. TypeSafe calls this model family System One. Official introduction

Imagine an inbox containing hundreds of customer messages. Before anyone writes a reply, several decisions need to happen: which team owns the issue, how urgent it seems, and whether a person should review it. That is the kind of work Jev is designed to help with.

Your application still decides what to do. A model’s recommendation to send a ticket to engineering does not itself create a ticket, contact a customer, or authorize a refund.

Explore the technical details

Three ways to ask a question

Primitive Use it when you need… An example question
Choice One answer from options you define Which queue should receive this message?
Score A rating against ordered descriptions How strongly does this article match our topic?
Noul The probability that a statement is true Does the customer explicitly request a callback?

You can ask several questions about the same input in one request. Jev evaluates them independently; your code combines the answers. Question types

The distinction matters: “How urgent is this?” needs a defined scale. “Does this mention a deadline?” asks a narrower yes-or-no question. Choosing the right question is part of designing the application.

What Jev does—and what you supply

Jev currently works with text, including text represented in JSON objects or arrays. It does not directly accept images, audio, or video, and it does not generate an email, essay, explanation, or code. System One documentation

You supply the relevant context, possible choices, scoring criteria, and rules for handling the result. For a document workflow, that could mean extracting text first, asking Jev to judge relevance, and using a separate tool to summarize the selected documents.

Start with practical use cases, or read Jev vs. LLMs to decide which parts of your workflow belong with which tool.

Explore the technical details

Why the probabilities matter

TypeSafe describes its training approach as Reinforcement Learning for Calibrated Decisions, or RLCD. The aim is for reported probabilities to match outcomes across many predictions. For example, among comparable predictions assigned a 70% probability, the event should occur about 70% of the time in a well-calibrated system. That is a statistical property, not a promise about one answer. TypeSafe’s AI primer

For a developer, this opens a useful design choice: act on clear cases, collect more context for ambiguous ones, and send difficult cases to a person. The thresholds still need evaluation on the task you actually care about.

Where to go next

  • Developers: start with one semantic check, then keep execution and permissions in code.
  • Support teams: separate queue selection, urgency, and the decision to involve a person.
  • Search and RAG teams: find relevant evidence before asking another model to write an answer.
  • Security teams: explore review signals while retaining established controls and analyst judgment.

The use-case guide follows these roles and includes a complete official document-search walkthrough. The limitations guide explains where the interface, evidence, and operational assumptions stop.

Common questions

Is Jev a chatbot?

No. Its output is constrained to the decisions you define. It can help a chat application choose a route or evaluate a response, while a separate model or template provides the actual wording.

Does structured output mean the answer is correct?

No. A response can fit the expected type and still choose the wrong category. Treat format guarantees and decision accuracy as separate properties. Our comparison guide explains the distinction.

Can I try it right now?

Start with our Playground: load an example or ask your own question. Live checks depend on the site’s availability and usage limits. For your own API integration or the official Playground, follow the getting-started guide and the official TypeSafe website; a TypeSafe account and the required product access are separate from this website.

What does it cost?

See the pricing guide and calculator for the published input-token rate, a practical estimate, and what the estimate leaves out.

Does this website’s language selector describe Jev’s language support?

No. The website’s 12 reading languages describe this guide’s translations. They are not a model capability list, an accuracy claim, or evidence that the same question works equally well in each language. Evaluate the language and terminology of your own workflow.

Is this the official Jev website?

No. What is Jev is an independent educational guide, unaffiliated with TypeSafe AI. We link to primary sources and label examples by their evidence. Read our editorial approach.

Sources & further reading

Official documentation and linked community reports support this guide. Community observations are attributed to their authors.

  1. TypeSafe introduction
  2. System One explained
  3. AI primer and calibration
  4. TypeSafe official website
How we check our sources
Get started

Try a small decision

Try your own text, question and rules with Jev. This site's free demo allows up to 10,000 characters in total.

Try it