Jev AI guide
What Is Jev AI? A Developer's Guide to Fast Decisions
Jev AI is TypeSafe AI's System One model: instead of generating text token by token, it evaluates your input against structured questions and returns typed decisions. Note the name collision: bare “what is jev” searches mostly return Japanese Encephalitis Virus medical pages. This guide is only about Jev AI, the developer model.
The official announcement describes Jev as reaching similar intelligence to existing LLMs on System One tasks while running far faster and cheaper (TypeSafe AI blog, September 2026). Treat vendor speed and cost multiples as vendor claims, not this site's measurements: we have made zero live calls. What you can verify today is the interface shape, documented at the official API reference.
System One decisions instead of chat
A chat model answers anything in prose. Jev answers only the questions you define, in types you choose. You send a state string plus questions with explicit criteria, and you get back a choice, a 0-1 noul value, or an ordered score. That constraint is the whole point: it turns open-ended generation into a reviewable decision with probabilities attached.
Three question types: choice, noul, score
choice · Select a category
Define named options and what each means. The documented response includes the selected option, probabilities and confidence.
Example: route a support ticket to billing, technical, sales, or other. The documented response includes the selected option, per-option probabilities, and a confidence value.
noul · Evaluate a proposition
Ask a yes/no question. The documented noul value is between 0 and 1; it is not a ready-made business decision threshold.
Example: does this request need human review? A 0-1 value is not a decision until you calibrate a threshold on your own labeled data and keep the human path open below it.
score · Use ordered criteria
Define ordered levels. The documented score uses zero-based levels and their probability-weighted value, not a percentage.
Example: grade urgency across ordered tiers. Scores are zero-based levels with a probability-weighted value, not percentages.
How developers actually use it: the request pattern
Send state, model, and questions to the TypeSafe evaluation endpoint, read the typed answers, and apply your own decision policy on top. The full copy-paste walkthrough with cURL, TypeScript, and error handling lives on the API example page; the support-ticket scenario shows six worked cases.
What a minimal request looks like
The smallest useful request names one question and four explicit criteria. State is the ticket text. Model is the requested model identifier; this site's examples use jev-latest, which the docs present as a movable alias rather than a pinned version. Questions carry the actual contract: each named option needs a one-line ownership definition, plus an other bucket for everything unclear, multi-team, or out of scope.
{
"state": "Please send my invoice for August.",
"model": "jev-latest",
"questions": {
"department": {
"type": "choice",
"instructions": "Route this support request to one team.",
"criteria": {
"billing": "Invoices, payments, refunds or existing charges",
"technical": "Bugs, errors, outages or integrations",
"sales": "Buying, demos, plans or presales questions",
"other": "Unclear, multi-team or out-of-scope requests"
}
}
}
}Start from this exact shape when you write your first local test. Change the criteria words only when your team ownership actually differs, and keep the other option even if you think your taxonomy is complete. Real queues always produce messages that fit no team, and other is where the review queue begins. Five de-identified tickets from your own work, run locally with your own key, teach more than any marketing benchmark. Pin a stable model version in your own code when one is available; jev-latest can drift and yesterday's labels may not reproduce on tomorrow's alias.
Jev API example with code · Ticket routing scenario · Evaluation method note
Jev vs general LLMs: decisions versus generation
A general LLM generates open-ended text and you parse intent out of prose. Jev inverts that: you declare the decision up front and the model returns a typed value with probabilities. The table below keeps vendor speed and cost multiples out because this site has not measured them; it compares only the interface contract you can verify in the docs.
| Aspect | Jev (decision model) | Chat LLM (generation model) |
|---|---|---|
| Output | choice, 0-1 noul, or ordered score | free-form text |
| Contract | questions and criteria fixed per request | prompt interpreted per call |
| Review path | probabilities point at the review queue | you build parsing plus validation yourself |
| Best fit | routing, triage, grading with a human fallback | drafting, summarising, open-ended help |
What Jev AI is not
It is not a chatbot, not a knowledge base, and not a replacement for human judgment on ambiguous or high-stakes requests. Structure is not evidence of accuracy: a well-formed response can still be wrong. Validate labels, calibrate thresholds, measure errors, and retain human review. This site publishes no live-verified fixtures or Jev performance results.
Structure is not evidence of accuracy.
A well-formed response can still be wrong. Validate labels, calibrate thresholds, measure errors and retain human review. This site has no live-verified fixtures or Jev performance results yet.
This is an independent, unofficial project, not affiliated with TypeSafe or the Jev team. We offer free tutorials and examples, plus a paid hosted Playground and Batch CSV evaluation that run the model on our servers via OpenRouter and are billed in credits. You can also export the code and run it with your own API key.
FAQ
Is Jev AI the same as JEV the virus? No. Same letters, unrelated topics. Search engines mix them, so always include “AI” when searching developer content.
Where is the official documentation? Start at the TypeSafe API reference, then come back for worked examples.
Does this site offer a Jev API? No. It offers guides and copy-paste code for your own key.
Copy the API example →