Jev AI: What Is Jev and How Does the Jev Model Work?
Learn what Jev AI is, how the Jev model returns typed choices, probabilities and scores, and where it fits in software and AI agent workflows.
Jev is a structured AI decision model from TypeSafe AI. Instead of generating a paragraph, the Jev model evaluates shared state against questions whose possible outputs are defined in advance. The response contains typed answers, probabilities, confidence and scores that software can use directly.
TypeSafe calls this model family System One Models. The design goal is fast judgment inside software, where a bounded answer is more useful than generated prose.
This guide is about the Jev AI model released by TypeSafe AI. It is not about the unrelated medical abbreviation JEV, which is commonly used for Japanese encephalitis virus.
What is Jev AI?
Jev AI behaves more like an intelligent decision function than a chatbot. An application supplies context plus explicit questions such as “Which queue should receive this ticket?”, “Does this action require approval?” or “How risky is this deployment?” Jev returns an answer constrained to the options or scale supplied by the application.
That narrower interface makes Jev useful at high-frequency decision points where open-ended text would need extra parsing, validation and error handling. It does not mean every Jev judgment is correct; it means the output shape and available answers are bounded before the call.
How the Jev model works
A Jev request separates two ideas:
- State contains the facts, evidence, or content being evaluated.
- Questions contain the judgments Jev should make about that state.
AutoJev supports the three Jev question types used by TypeSafe workflows and the configured Decisions API:
- Choice selects one option from a set of named criteria.
- Noul returns a probability for a yes-or-unknown judgment.
- Score places the state on a rubric defined by the caller.
Several questions can evaluate the same state in one request. A tool-call review can ask for an action verdict, a risk score, and a human-confirmation probability together.
Jev compared with a generative LLM
Jev and a generative language model solve different parts of an agent workflow.
| Need | Jev | Generative LLM |
|---|---|---|
| Pick from known actions | Strong fit | Possible, but requires output parsing and validation |
| Return calibrated probabilities | Native response | Depends on prompting and provider behavior |
| Write code or prose | Not supported | Strong fit |
| Explain a complex plan | Not supported | Strong fit |
| Route, classify, score, or gate | Primary use case | Often slower and more expensive for repeated decisions |
A practical agent stack uses a generative model to plan and create, then calls Jev at bounded decision points. The Jev answer does not replace deterministic checks, permissions, budgets, or human approval.
Good Jev use cases
Jev works best when the possible actions are known before the request:
- Route a task to one of several AI models.
- Classify a support request or operational event.
- Score release, compliance, or tool-call risk.
- Decide whether research evidence is sufficient.
- Check whether an agent task is complete.
- Estimate whether a human needs to review the next action.
- Choose whether a workflow should proceed, pause, split, or stop.
When not to use Jev
Do not use Jev when the task requires it to write an answer, generate code, create tool arguments, inspect files, or explain its reasoning in prose. Those jobs belong to a generative model or deterministic application logic.
Typed output prevents invalid response shapes, but it does not guarantee that every decision is semantically correct. Test important decisions against your own data, keep irreversible actions behind normal approval boundaries, and combine model judgments with deterministic policy checks.
How AutoJev uses Jev
AutoJev turns the Jev decision model into an operational layer for AI agents:
- A remote Jev MCP server exposes typed tools to MCP clients.
- REST presets define stable input contracts for common decisions.
- Portable Skills teach agents when to call each decision tool.
- Deterministic guidance maps model outputs to workflow actions.
- Bearer authentication keeps the upstream provider key on the server.
- The AutoJev Playground lets you test requests without writing a client.
AutoJev does not replace Jev and does not hide the underlying decision model. It packages Jev for agent workflows that need repeatable inputs, discoverable tools, and explicit action policies.
Jev AI questions people ask
Is Jev a chatbot?
No. Jev does not produce open-ended chat responses. It evaluates the state and questions supplied by software, then returns typed decisions and probabilities.
Is Jev a replacement for an LLM?
Not for generation. A generative LLM remains the better tool for writing, coding, planning and explanation. Jev can complement that model by handling classification, routing, scoring and verification steps.
Can Jev make a wrong decision?
Yes. A typed response guarantees a valid interface, not a correct judgment. Production systems should calibrate thresholds on their own data, preserve deterministic validation and escalate uncertain or high-impact decisions.
Explore the Jev topic cluster
- Jev AI for agents explains the agent architecture and practical decision patterns.
- Jev model guide covers state, Choice, Noul, Score and response handling.
- Jev MCP server shows how to connect Codex, Claude Code, Cursor and other MCP clients.
- Jev Skills explains the umbrella router and five focused AutoJev workflow Skills.
- Jev online playground lets you inspect real request and response shapes.
Official Jev resources
- TypeSafe System One documentation
- TypeSafe introduction to Jev
- Jev on OpenRouter
- Jev on Cloudflare Workers AI
Continue with the AutoJev integration guide or configure the Jev MCP server.