Choice, Score or Noul?

Start with the answer your code needs

A department name, an urgency level and a cancellation flag are different outputs. Trying to hide all three in one broad question makes the result harder to inspect. Choose a primitive for each judgment, then combine the answers in code.

The official primitive reference defines Choice, Score and Noul. All can evaluate the same supplied state, but they serve different purposes.

Choice: select from a defined set

Use Choice when the output is a member of a list: billing, technical support or another department. Write a description for each option so the boundaries are clear. If the list may be incomplete, include an explicit fallback category.

Choice returns the selected option, probabilities across the options and a confidence value. See the Choice specification.

For a model router, first remove models that cannot handle the required media or context size. Then ask Jev to choose among the remaining options. Capability checks belong in code; the semantic preference is the judgment.

Score: evaluate ordered levels

Use Score when you can describe a meaningful progression. A relevance rubric could distinguish unrelated, tangential, useful and directly answering the question. The descriptions do more work than bare numbers.

Score includes a value, a legend for its levels, probabilities and confidence. Its value can fall between levels. It is not a substitute for measuring an exact quantity. See the Score specification.

A practical exercise is to have two people apply the rubric to the same examples. If they cannot agree on what separates the levels, clarify the rubric before judging model quality.

Noul: evaluate one yes/no condition

Use Noul for a specific proposition, such as whether a message explicitly requests a refund. Its noul value represents the probability of yes. A value near the middle expresses uncertainty, not a medium amount of the attribute.

Noul does not include a separate confidence field. See the Noul specification. Decide how your application handles uncertain probabilities rather than automatically rounding every value into a decision.

A quick selection table

These are design examples, not returned model values. The same business process can use all three without making them a single question.

Write the full criterion

Question identifiers are for your code. Put the complete judgment in the instructions instead of relying on a key such as urgent to explain what urgent means. Clear criteria are especially important for edge cases.

Next, read how to interpret confidence or see the field-mapping example.