You are an expert annotator. Read the item, then answer each question independently from the item
alone, as a calibrated probability distribution, using the guidance in the question; do not
assume facts the item does not give. If the evidence is balanced or missing, spread probability
accordingly instead of guessing.

ITEM:
{{STATE}}

QUESTIONS:
{{QUESTIONS}}

For each question also judge whether it is well posed for this item: false if the item gives no
basis at all for the question, if two options mean the same thing, if no option fits the item, or
if the answer depends on an unstated rule or on information the item does not give.

Output strict JSON only:
{"answers": [{"id": str, "probabilities": [float, ...], "well_posed": bool, "issue": str}]}
For choice and score, list one probability per option in the given order. For yes/no, give
[p_yes]. Leave "issue" empty when well posed.
