jev choice vs noul

The Difference Is the Meaning of the Answer Space

A Choice asks a relative question: which supplied option best matches the state? Its option probabilities form one distribution, so increasing support for one option necessarily changes the others. A Noul asks whether one proposition holds. Separate Nouls do not compete and can all be high, all be low or disagree in ways that require investigation.

Choose the primitive from the real-world label structure, not the desired UI. Radio buttons do not make a taxonomy complete, and checkboxes do not prove labels are independent. Write down whether exactly one, at most one, several or no labels may be true before designing the request.

One Support Ticket Modeled Two Ways

“I was charged twice after the export failed, and payroll is blocked” legitimately activates several Nouls, while a downstream owner field may still require one Choice. Trying to encode every signal in one queue label loses information. Trying to route with unrelated Nouls pushes conflict resolution into accidental threshold order.

Choice Needs an Explicit Escape Hatch

A Choice selects from the criteria supplied. If the real answer is missing, the winner is merely the least-wrong listed option. Include other only when it has a clear operational destination; a vague catch-all can absorb taxonomy errors without revealing them. For high-risk routing, pair the Choice with a Noul such as “the selected handler is qualified for this request” and require both tests to pass.

Log no-match and low-fit volume as product feedback. Repeated cases may justify a new category, better upstream filtering or a different system. Do not continuously expand a candidate list without retesting because more overlapping criteria can reduce separation.

Equivalent-Looking Questions Need Not Produce Equivalent Numbers

Do not expect the probability of billing inside a four-way Choice to equal a separate Noul asking whether the message is about billing. The contexts differ: one compares billing with three alternatives, while the other compares yes with no. TypeSafe’s Jev 1.13 notes explicitly warn against assumed arithmetic invariants across separately formulated questions.

Likewise, a positive Noul and a separately phrased negative Noul are not complements that must sum to one. If an application needs a complementary distribution, use one modeled question and derive the complement in code. Validate the exact formulation on labels through the calibration procedure.

Evaluate the Primitive and the Resulting Decision

The broader primitive guide covers response fields and confidence versus probability explains distribution concentration. This page’s narrower conclusion is structural: mutually exclusive alternatives and independent propositions are different statistical objects.

• For Choice, report the confusion matrix, per-class precision/recall, no-match rate and distribution calibration.
• For Nouls, report each condition separately, including base rate, calibration and threshold coverage.
• Measure co-occurrence and contradictory outcomes for multi-label designs.
• Compare downstream routing or action quality, not only model-label agreement.
• Replay after taxonomy, wording, state projection or model changes.