TypeSafe documents three primitives for Jev. Each is a different shape of judgment. Pick the shape before writing a prompt: the type determines what your code can expect back. The primitives reference has the current wire format and detailed examples.

Choice: select among named options

Use Choice when the application has a known set of destinations or labels. For a support ticket, define billing, technical, and sales with short, distinct descriptions. The response includes the selected choice, probabilities for options, and confidence. Add an explicit uncertain or other option when your real inputs can fall outside the list. Do not make categories overlap unless your application has a rule for resolving that overlap.

Score: use an ordered rubric

Use Score when the answer is a level on a descriptive scale, such as calm, frustrated, or very angry. Write what each level means in observable terms. A score is not an exact measurement of an unstated numeric quantity. The jev-1.13 limitations specifically caution against treating score levels as a precise number between rubric points.

Noul: evaluate a yes/no proposition

Use Noul for a narrow proposition such as “Does this message explicitly request a refund?” Its result is a value from 0 to 1. State the proposition precisely and keep missing evidence distinct from evidence against it when your workflow needs that distinction. If you need a reason in natural language, use another system or show the source text; Jev is not a text generator.

Try one state, several questions

The API can evaluate multiple independent questions against the same state in a single request. A ticket can be routed with Choice, marked urgent with Noul, and rated for frustration with Score. The questions are evaluated independently, so your code must combine their outputs and enforce any business constraints. Start with one question, check it on examples, then add only questions that change an action.

Jev Browser Automation shows the Choice shape in a browser loop: the model selects one action from the elements available on the page, while code owns retries and stopping. Jev Gates Agent Memory Saves shows a different bounded judgment, checking a note against its source conversation. Neither listing is a substitute for testing your own options and thresholds.

Review your design

Can a human answer each question from the state supplied? Are the options complete and distinct? Does a wrong answer have an acceptable fallback? If a question contains several judgments joined by “and,” split it. Continue with state and question design and confidence-based review.