Choosing between Choice, Score and Noul for a decision passed to a decision model

Goal

Express a judgment so that the model answers exactly the question the code needs, with an answer type the code can branch on without interpretation.

Prerequisites

A closed answer space. If the answer is free text, a number to be computed or a date to be compared, the decision model is the wrong tool for that part (see the documented failure modes). The API reference fixes the shapes: a Choice needs a map of option to description with at most 255 options; a Score needs an ordered array of at least two and at most ten level descriptions; a Noul takes optional descriptions of what yes and no mean.

Steps

1. Write the judgment as one sentence. If the sentence contains "and" joining two independent properties, split it into two questions; the documentation states that questions are evaluated independently, so adding a question does not change another question's answer.
2. Decide the answer type. One of several handlers, categories or candidates: Choice. A position on an ordered scale (severity, urgency, quality): Score. A single property that is present or absent: Noul.
3. For a Choice, add an explicit option for "none of these" or "not stated" whenever the state may not contain an answer. The model always returns the highest-probability option; without an abstain option a bad fit is hidden inside a low confidence.
4. For a Score, describe every level in concrete words that a person could apply, in increasing order. Do not plan to interpolate an exact magnitude between levels; the jaggedness page says score levels are weak in numerical calibration and that the expectation is fit for a threshold check only.
5. For a Noul, phrase the question so that yes is the natural reading of the words. The documentation reports worse results when the true criterion maps to "no", and a Noul carries no confidence value, so decide in advance which probability counts as yes.
6. Put the material the question refers to into the state under a descriptive key and name that key in backticks in the instructions, so the model reads the intended part.
7. Give each question a key that says what it decides; the API states that keys are not sent to the model.

Expected result

A request whose questions are atomic, whose options cover the cases including "none", and whose answers map one-to-one onto branches in code. Reading the probabilities shows a peaked distribution on clear inputs and a flat one on ambiguous ones.

Limits and test basis

This is a design procedure derived from the vendor's API reference and stated failure modes; no accuracy measurement is claimed. Thresholds and the abstain option must be tested on the caller's own labelled examples before an automated action depends on them.