Jev Noul, Choice, and Score: Picking the Right Answer Shape
Noul, Choice, and Score define three different decision contracts. Pick the one that matches the next line of code, then calibrate it as its own instrument.
Design from the branch backward
A primitive is not a formatting preference. It determines the answer space, the probabilities you receive, and the mistakes the application can make. Begin with the operation that consumes the result. If code must select one queue from a closed list, the natural contract is a Choice. If it compares an ordered severity against a threshold, use a Score. If it needs evidence for one proposition, use a Noul. The closer the model output is to the actual branch, the less hidden translation logic sits between inference and behavior.
Do not begin with a vague business question and choose whichever primitive makes the demo easiest. “Is this a good customer?” could mean renewal risk, contract value, product fit, or policy eligibility. Separate those meanings first. A bounded model can make one fuzzy judgment, but it should not silently decide which definition matters. Product policy belongs in named criteria and visible code.
Use Noul for one proposition
Noul answers the probability that a statement is true. It fits signals such as “the message explicitly asks to cancel” or “the supplied passage supports this claim.” The value near either edge indicates directional evidence; a value near the midpoint indicates that the state does not clearly favor true or false. That midpoint is uncertainty, not a middle amount of the underlying concept. A question about how severe something is should not be squeezed into a Boolean probability.
Write the proposition so a human labeler could apply it consistently. Define scope, time boundaries, and what to do when the evidence is absent. Then decide how the probability enters the system. You may use a high threshold for automation, retain a lower threshold for prioritization, and send the middle to review. Those cutoffs are operational policy and must be chosen from labelled examples rather than copied from another Noul or another domain.
Use Choice for relative selection
Choice returns one member of a declared set along with a distribution over that set and a confidence value. It is well suited to routing, document type, intent, or any case where downstream code already switches over a finite enum. The criteria should make alternatives distinct. If two options overlap, the model is being asked to resolve a taxonomy problem that the product team has not resolved.
Include an escape option when real inputs can fall outside the list. Without “other,” “unknown,” or an equivalent route, the distribution still has to allocate its mass among the available categories. A neat winning label can therefore conceal an incomplete taxonomy. Review the full distribution rather than only the winner: a narrow margin between the top two options can justify escalation even when the returned category is structurally valid.
Use Score for an ordered rubric
Score operates over levels you define in order. This suits concepts such as urgency, evidence strength, or support impact when each level has a concrete description. Levels should describe observable distinctions, not decorative labels such as low, medium, and high with no criteria. A reviewer should be able to explain which evidence moves an example from one level to the next.
The returned position can lie between levels because it summarizes a probability distribution, but the scale is ordinal rather than a physical measurement. A position between “minor inconvenience” and “blocked workflow” is not a percentage of inconvenience, and arithmetic on it can imply precision that the rubric never defined. Thresholding or ranking may be appropriate after validation; multiplying it by revenue or presenting it as a measured magnitude is not.
Do not translate thresholds between primitives
The same natural-language idea can behave differently when expressed as a Noul, a two-option Choice, or a Score. A Choice asks which option wins relative to the others. A Noul asks for an absolute belief in one proposition. Several Nouls are not required to sum to one, while Choice probabilities describe one closed competition. This is why a threshold tuned for one form cannot be carried into another just because both appear to answer yes or no.
Treat each primitive and question revision as a separate statistical instrument. Build an evaluation set, keep the raw distributions, and measure errors by consequence. If you change from Choice to Noul, reset the calibration work. If you add a Choice option, rerun the test because every existing option now competes in a different set. Schema compatibility is not behavioral compatibility.
Compose multiple judgments in one request
A practical request can mix all three shapes against one shared state. A ticket workflow might use Choice for destination, Score for customer impact, and Noul for explicit cancellation intent. Code can combine them: a likely cancellation with high impact may enter a retention review, while an uncertain destination stays in general triage. Each model output remains small and named, and the final route is a deterministic expression the team can test.
Parallel questions do not create a hidden reasoning chain. If one judgment truly depends on the output of another, make that dependency an explicit second stage or redesign the inputs. Also verify that each question can be answered from the supplied state. More questions are not automatically more evidence. The best set is the smallest set of independently useful signals that changes a real branch or improves an evaluation.
What to carry into your next build
- Choose the primitive by the output contract your next code path needs.
- A Noul midpoint means uncertainty; a Score represents an ordered rubric; a Choice is relative to its option set.
- Provide an escape route for incomplete Choice taxonomies.
- Calibrate every primitive and question revision independently.
Editorial method and source note
This article is an original synthesis based on Learn Jev's “Noul, Choice and Score,” checked against TypeSafe's official introduction and primitive documentation. The independent guide usefully surfaces edge cases, but its examples are not a substitute for current API docs or evaluation on your data. No source outline, prose, image, or code is reproduced.
Facts and product details should be checked against the linked source and current official documentation before making production decisions.