Plain-language guide
What is Jev?
Jev is TypeSafe AI’s model for making small, typed judgments inside software. It receives application state and questions with predefined answer shapes, then returns values, probabilities, and confidence for code to use.
The document, event, record, or context being judged.
A narrow Choice, Score, or Noul with explicit criteria.
A typed answer with probabilities and confidence.
Thresholds, routing, fallback, and exact computation.
Why it is different from a chat model
Most language models generate a string one token at a time. Applications then ask them for JSON, parse the result, validate the shape, and decide whether the text can be trusted. Jev gives up free-form text generation. TypeSafe positions it as a model for fast, structured decisions that software can consume directly.
The company calls this category “System One models.” That is TypeSafe’s terminology, not a neutral standard adopted across the industry. Jev Directory uses Jev for the specific model and “typed decision model” as a plain description.
The three question types
Choice selects from options defined by the developer and returns the selected choice, probabilities over the options, and confidence. It works for bounded classifications such as issue type or destination. Include an explicit no-match option when the real world can fall outside the list.
Score evaluates a state against an ordered rubric. The criteria describe what each level means, so application code can rank or combine the result. Noul asks whether one statement is true and returns a value from zero to one. It is useful when several independent properties may all apply.
Keep questions atomic
TypeSafe recommends asking one well-scoped thing at a time. “Is this support request urgent?” is easier to inspect than “How should we handle this customer?” If a decision depends on urgency, account risk, and topic, ask those separately and combine the outputs in code.
This separation keeps product policy visible. A developer can change a weight, threshold, or fallback without hiding the business rule inside rewritten prose. It also makes evaluation more useful because each question can fail independently.
Where Jev fits well
Good candidates have unstructured context but a bounded next decision: routing a request, checking whether a policy applies, scoring a rubric, filtering agent context, or selecting items for human review. The output should lead into deterministic code rather than become user-facing prose.
Jev is a poor fit when the product needs an explanation, long-form synthesis, original text, or a multi-step solution from the model itself. A typed output also does not guarantee a correct judgment. Representative data, clear criteria, appropriate thresholds, and a fallback path still matter.
A small routing example
Imagine an inbox with three handlers: billing, technical support, and human review. The message is the state. A Choice question defines billing, technical, and unclear. Jev returns probabilities for those options. Code sends a confident billing answer to the billing workflow, while an unclear or low-confidence result goes to a person. The model proposes the route; application policy decides whether to act.
Start with an existing pattern
Begin with intent routing for a bounded destination, confidence-gated routing when consequences differ, or composite scoring when several dimensions need explicit weights. For implementation syntax and current model identifiers, use the official TypeSafe documentation.
Jev Directory is independent of TypeSafe AI. We separate vendor documentation, community claims, and behavior we have actually tested; the evidence note on every entry says which one you are reading.