Noul, Choice and Score: choose the answer shape

Choose the shape that expresses the question you can actually define. TypeSafe documents three primitives for Jev: Noul, Choice and Score. They constrain the answer form; they do not establish correctness. Primary primitive reference.

Jev accuracy in our use is unmeasured because the evaluation was dropped. Everything below describes the contract or a fictional, unexecuted example.

Noul: one binary proposition, a probability

Noul returns a number between zero and one representing the probability that the supplied proposition is true. The native result is a probability, not a literal boolean. The consumer decides how to interpret it.

An illustrative question could be: “Does this request explicitly ask for a meeting?” Define what counts as an explicit request, including negation and quoted text. Do not treat a large probability as permission to book a meeting.

If the application needs a boolean, its own threshold and review rule turn the probability into that disposition. This guide supplies no validated threshold. A policy must also decide what to do with missing context or uncertainty.

Choice: one supplied option, with a distribution

Choice returns a selected option, probabilities across the supplied options and a confidence value. Write option descriptions that distinguish the outcomes.

For a fictional queue, use destinations such as “scheduling review”, “document request review” and “other or unclear”. These are invented design labels, not labels from a tested dataset. Including a review option does not prove the model will select it when appropriate.

A Choice question supplies the menu. The consumer still needs to check whether the returned destination is permitted and whether the item should move at all.

Score: ordered described levels, capped at ten

Score uses ordered descriptions and returns a probability-weighted mean of their positions, a distribution and confidence. It can fall between levels. The current native contract allows at most ten described levels.

Define one dimension. A fictional completeness rubric might distinguish “no relevant detail”, “some requested detail” and “all requested fields supplied”. That is an illustration of ordered criteria, not a tested rubric or a statement that Jev checks every field reliably.

A native criterion list containing every integer from zero through ten has eleven levels and exceeds the documented cap. If application code maps a smaller rubric onto another display scale, record the mapping separately; it adds no native level or evidence of precision.

A mean can hide different distributions. Inspect the returned probabilities when the distinction matters instead of assuming the same score implies the same uncertainty.

Confidence is not an accuracy measurement

TypeSafe describes Choice and Score confidence as a statistic derived from the returned probability distribution. Noul has no separate native confidence field. Concentration within a model’s answers is different from agreement with independently labelled outcomes in your task.

No confidence threshold in this guide has been evaluated. Decide the review and failure requirements before any future test, then keep the result within the authority boundary described by the canonical authority guide.

Write the consumer contract

Record the primitive, exact question, criteria and version. Then define what the application may do with the result, when it holds for review and which existing rule takes precedence. Those are original design recommendations.

If the request needs a document drafted, none of these answer shapes supplies that document. If the next step needs a defined label or position, the interface is a candidate worth specifying, with suitability still unverified. Return to bounded decisions for the wider decision.

Sources and limits

Primary documentation checked 7 October 2026: primitives, Noul, Choice, Score and confidence. These pages support the documented contract, not a separate primitive launch date or local performance result.