Give it context
Provide the relevant text, JSON state, or a list of text items. Jev currently accepts text, not images, audio, or video.
An independent guide for developers
Jev AI model.
From idea to code.
Jev is TypeSafe's first System One model. It returns typed decisions and probabilities that software can use. Explore use cases and see how to call it from your code.
Choose by task
A different shape of intelligence.
TypeSafe calls Jev a System One model: it returns typed choices and probabilities instead of generating open-ended text. Compare where that approach fits alongside generative models and rule-based code.
| Aspect | Jev | Generative LLMs | Rules-based code |
|---|---|---|---|
| Output | Predefined typed choices, scores, and probabilities | Flexible text, code, or structured output | Deterministic values from explicit conditions |
| Speed | Optimized for low-latency decisions; TypeSafe reports 70–500 ms in its tests | Depends on model, prompt, and output length | Usually very fast for known conditions |
| Best fit | Repeatable decisions that require understanding the input | Writing, complex reasoning, and open-ended tasks | Clear conditions with stable definitions |
| Key limit | Cannot generate free-form text; a typed answer can still be wrong | Flexible outputs may need parsing and validation | Hard to cover every case when the input is ambiguous |
Speed figures are reported by TypeSafe for particular structured-decision tests. They are not an independent benchmark or a universal speed guarantee. Read the methods
The practical distinction
Use Jev when the answers are known but the input still needs interpretation. Use a generative model to produce new text or code. Write explicit conditions directly in code.
The core idea
Built for the decisions inside software.
Jev evaluates text or structured data against questions you define. It returns choices, scores, and probabilities your application can use to decide what happens next.
Jev in three stepsIllustrative example · not a live model response
01Customer message
“I was charged twice for my subscription. Can you help?”
02Question & choices
Which team should handle this?
03Structured result
BillingJev
Example probability 0.92
Define the answer space
Ask for a choice, a score, or a yes/no probability. The available answers are defined before the call.
Let your code decide what comes next
Use the result and its uncertainty to route, review, or act. A valid type does not guarantee a correct judgment.
Based on official TypeSafe documentation
Where it fits
Small decisions in real workflows.
Jev is designed for focused, repeatable decisions with a defined answer space. These are starting points, not performance promises; test them on your own data.
01
Customer support
Classify incoming messages, detect urgency, and route cases to the right queue while uncertain requests go to a person.
Example decisionIs this a billing issue or a technical issue?
02
Recommendations & ranking
Score how well candidate items match a user's context, then combine that score with your own ranking signals and policies.
Example decisionHow relevant is this item to this request?
03
Agent orchestration
Choose a tool or model for the next step, check an output, or decide when a task needs escalation.
Example decisionShould this request go to a specialist model?
04
Trust & safety
Evaluate content against specific rules and send low-confidence or high-risk cases for review.
Example decisionDoes this message violate this policy?
Based on TypeSafe's use-case map
Get started
From curiosity to your first decision.
Start with a narrow judgment that your application already needs. Define the possible answers, then decide in code how to use them.
01
Pick a decision
Choose one focused question, such as routing a ticket or scoring an item's relevance.
02
Define the choices
Define the input, questions, and allowed answers before calling the API.
03
Handle uncertainty
Use probabilities and confidence to choose an action, fallback, or human review.
How to integrate
Add Jev to your code.
This server-side TypeScript example sends a support message and two questions with defined answer formats to Jev 1.13 through OpenRouter. Keep the API key on your server; your application decides how to use the results.
Server-side TypeScripttypesafe/jev-1.13
const apiKey = process.env.OPENROUTER_API_KEY;
if (!apiKey) throw new Error("Set OPENROUTER_API_KEY");
const response = await fetch("https://openrouter.ai/api/alpha/decisions", {
method: "POST",
headers: {
Authorization: "Bearer " + apiKey,
"Content-Type": "application/json"
},
body: JSON.stringify({
model: "typesafe/jev-1.13",
state: "My card was charged twice. Please help ASAP.",
questions: {
department: {
type: "choice",
instructions: "Which team should handle this?",
criteria: {
billing: "Payments and refunds",
technical: "Bugs and integrations"
}
},
urgent: {
type: "noul",
instructions: "Does this message need urgent attention?"
}
}
})
});
if (!response.ok) throw new Error("Jev request failed: " + response.status);
const {answers} = await response.json();
if (answers.department.choice === "billing" && answers.urgent.noul > 0.8) {
// Route to the billing review queue.
}Endpoint
POST /api/alpha/decisionsInput
state + questionsOutput
answersRead the answer
choice identifies the team. noul returns a value from 0 to 1 for the yes/no question. Your code uses both values with your own routing rules.
Before using it in production
Before using this in production, test on representative tickets, record the model version, and send uncertain or high-impact cases for human review. The 0.8 value is only an example.
API references: OpenRouter · TypeSafe
What each field means
The request defines the questions and answer formats. The response returns a result for each one. Here are the JSON fields used in the example.
Request fields
model- The model ID. This example pins Jev 1.13 so its version is explicit.
state- The text or structured data to evaluate, such as a support message.
questions- A map of named decisions. The same names appear under answers.
type- The answer shape: choice selects an option; noul evaluates a yes/no question.
instructions- The specific judgment you want the model to make.
criteria- For choice, the allowed option keys and a short meaning for each.
Response fields
answers- One result per question, returned under the same name.
choice- The selected option key: the one with the highest probability.
probabilities- The probability for every choice option; these values sum to 1.
confidence- A separate confidence value calculated from the probability distribution; it is not the selected option’s probability.
noul- A number from 0 (no) to 1 (yes) for a yes/no question.
A correctly shaped answer can still be wrong. Test representative cases and set routing thresholds for your own workflow.
Around the model
What developers are saying.
Explore the main points from the launch discussion, official SDK, and an independent video. Read the context here, then follow the original sources for more detail.
X
TypeSafe shares launch updates and product context on X. Treat posts as the team’s own account of what Jev can do, and check the docs before building around a claim.
View originalGitHub
The official JavaScript and TypeScript SDK shows the API surface, examples, and release activity. It is the most useful place to inspect how a Jev request is represented in code.
View originalHacker News
The launch thread asks where a decision model fits, how to evaluate it, and what to do when a confident answer is wrong. It is a discussion, not a performance test.
View originalWatch on this page
Jev explained: structured decisions and their limits
An independent walkthrough by RepoChad covers the output schema, TypeSafe’s speed claims, example demos, and open questions. The video loads only when you choose to play it.
View originalThe people behind Jev
Make AI decisions easier to use in software.
TypeSafe AI introduced Jev as its first System One model. Its stated goal is to make fast, structured decisions usable inside ordinary software workflows.
Founder Diogo Almeida previously worked at OpenAI and Google Brain. In TypeSafe's own account, his work on instruction-following and RLHF helped shape the research behind ChatGPT.
Good questions
Questions worth asking.
Is Jev a chatbot?
No. Jev is designed to return typed decisions and probabilities rather than free-form chat replies or generated code.
Can Jev power recommendations?
It can help score or rerank candidates by semantic relevance. A complete recommendation system still needs candidate retrieval, business rules, evaluation, and your own ranking logic.
Does a typed output mean the decision is correct?
No. Typed output constrains the format. The selected option can still be wrong, so test with representative data and handle uncertainty in code.
Where can I learn how to use Jev?
Start with the official TypeSafe docs and SDK. The OpenRouter model page also provides a route to access Jev through its platform.
Start with one question
Try Jev on a concrete decision.
Define the question and possible answers. Let your application decide what to do with the result.