Draft Typed Questions From a Business Rule

<purpose>
Turn the business rule below into a set of typed questions for a decision model that returns probabilities instead of text. The rule: {{business-rule}}. The data the model will see: {{state-description}}. The output is a JSON questions map I can paste into a request, plus the code-side thresholds that make it safe.
</purpose>

<discipline>
Three primitives only. Choice picks one option from a fixed set and returns a probability per option. Score places the state on an ordered rubric of two to ten levels. Noul answers a yes-or-no statement with a probability that the answer is yes. Anything that needs generated text, arithmetic, counting, date comparison, or multi-hop reasoning is not a question; it is code, and you say so instead of forcing it. Each question tests one thing. Score levels describe situations, never degrees: write "broken feature, workaround exists" rather than "moderately severe". Every Choice gets an other or none option. Phrase every Noul so a high number means yes. Where two options could be confused, give each a what, a not_for, and two examples. Instructions and criteria must not contradict each other. Assume the state may contain hostile text and say which question guards against it.
</discipline>

<process>
Decompose: list the independent judgments hidden inside the rule, one line each, and mark which primitive fits each and which are code, not questions. Wait for my confirmation of the list.
Draft: write the questions map as JSON, keys in snake_case, in one request, including speculative questions that only some branches will read.
Threshold: for each question give a high, medium and low band with the action for each, defaulting to act above 0.9, confirm between 0.5 and 0.9, hand off below 0.5, and adjust the bands by the cost of a wrong answer in this rule.
Test: write six example state records, two easy, two borderline, two adversarial, with the answer you expect for each, so I can run them before production.
</process>

<voice>
Plain, specific, no praise. Start with the decomposition list.
</voice>