TL;DR

  • Jev evaluates supplied state and returns bounded, typed decisions instead of generating prose.
  • Its three primitives are Choice, Score, and Noul.
  • Closed outputs prevent out-of-schema answers, but Jev can still make the wrong in-schema decision.
  • The best fit is a narrow semantic judgment that application code can verify, threshold, and act on.

Jev is TypeSafe AI’s model for decisions that software needs to consume. Give it application state and a set of typed questions, and it returns bounded answers with probability information. It does not write a paragraph that your code then has to parse.

TypeSafe calls Jev its first System One model. The name describes a model built for fast judgments such as classification, scoring, routing, or verification. It is a different interface from a conventional large language model, not a replacement for every task an LLM handles.

What does Jev return?

The API accepts a state, a model name, and a map of questions. The state can be a string, object, or array. Each question uses one of three primitives.

Primitive Question it answers Result
Choice Which allowed option fits best? selected option, full distribution, confidence
Score Where does this sit on an ordered rubric? expected score, level distribution, confidence
Noul Is this proposition true? probability from 0 to 1

A support system could pass in a ticket, use Choice to select a department, Score to rate urgency, and Noul to estimate whether the customer is asking for a refund. The TypeSafe introduction says those independent questions run in parallel against the same state.

That independence matters. One answer does not silently alter another question in the same call. If a later decision genuinely depends on an earlier result, the application should make another request with the relevant state.

Why does TypeSafe call it a System One model?

The product borrows the System One label for quick, focused judgments. In TypeSafe’s framing, a good question is something a knowledgeable person could decide in a few seconds once given the right context.

Examples include:

  • Which queue should receive this request?
  • Does this passage support the cited claim?
  • How severe is this incident on a defined rubric?
  • Which candidate is most relevant to this query?

Tasks that require extended reasoning, calculations, research, or an explanation do not fit the same shape. TypeSafe’s known limitations for Jev 1.13 explicitly say to keep arithmetic, counting, and date comparison in code. Open-ended generation belongs with a generative model.

How is Jev different from structured LLM output?

A conventional LLM can be prompted to return JSON or constrained to a schema. Jev starts from a smaller answer space: Choice picks from options supplied by the developer, Score uses ordered levels, and Noul returns a yes probability.

That removes a class of integration work. The application does not need to recover a category from prose or reject a newly invented label. It also exposes the complete probability distribution for Choice and Score, which lets code inspect ambiguity instead of accepting only the winning answer.

The boundary needs careful wording. A closed answer space stops Jev from inventing an out-of-schema category. It does not guarantee that the selected category is correct. Type safety solves the shape of the answer, not the truth of every judgment.

For a direct comparison of the two interfaces, read Jev vs. LLMs.

Where does Jev fit in an application?

The cleanest architecture keeps control flow in ordinary code.

  1. Code gathers the relevant state.
  2. Jev makes one or more narrow semantic judgments.
  3. Code reads the answer, probabilities, and confidence.
  4. Business rules decide whether to act, request confirmation, use another model, or send the case to a person.
  5. The application logs the actual model version and outcome for evaluation.

TypeSafe’s building guide recommends decomposing broad policies into atomic questions. Instead of asking Jev to decide whether a customer deserves a retention offer, a product could separately evaluate intent, frustration, eligibility context, and risk. Code would then apply the actual offer policy.

This split makes the policy inspectable. Product and engineering teams can change a weight or threshold without rewriting a long prompt that mixes judgment with business logic.

What is Jev good at?

The current use-case map emphasizes high-volume, bounded decisions:

  • intent classification and support routing
  • model, tool, or skill selection for agents
  • input, output, and tool-call guardrails
  • RAG passage filtering and reranking
  • citation checks
  • document classification and entity matching
  • semantic features for downstream predictive models

These are tasks where software needs a value it can branch on. A separate LLM can still generate the customer-facing response after Jev decides which path the request should take.

What should teams test before production?

Start with representative examples from the real workflow. Label the outcome you want, run Jev, and measure errors by category and consequence. A single aggregate accuracy number can hide the cases that matter most.

Then choose action thresholds by risk. Showing the wrong help article is recoverable. Blocking a customer, moving money, or making a compliance decision needs a higher bar and often explicit review.

TypeSafe’s confidence documentation also separates confidence from certainty. Confidence summarizes how concentrated a Choice or Score distribution is. It is useful for routing, but it is not a guarantee that a particular answer is correct.

Finally, pin the model version after tuning. The models page notes that jev-latest can move when a new stable release ships. The response reports the versioned model ID, so production logs can show which model made each decision.

The useful mental model

Jev is best understood as a semantic decision layer inside normal software. It handles the part that is hard to express as an exact rule, then hands a bounded result back to code.

Use an LLM when the output should be language, code, a plan, or an explanation. Use deterministic code when the rule is already exact. Jev fits between them when the application needs a quick judgment and already knows the allowed shape of the answer.

Sources

FAQ

What is Jev?
Jev is TypeSafe AI's first System One model. It evaluates structured or unstructured state against typed questions and returns choices, scores, probabilities, and confidence for software to use.
Is Jev an LLM?
TypeSafe presents Jev as a System One model rather than a text-generating LLM. It is designed for bounded decisions, not open-ended writing or conversation.
Can Jev hallucinate?
Jev cannot return an option outside the answer space defined by the developer, but it can still choose an incorrect allowed option. Production systems still need evaluation, thresholds, and fallbacks.