TypeSafe的决策模型：给一段内容和一组带类型的问题，返回概率、选项和打分，每个都带置信度；只按输入token计费。
state*
状态 · 所有问题都针对这段内容：文字，或一段JSON。英文最准。
默认 "I was charged twice for my subscription this month. Please refund the extra charge."
questions*
问题 · JSON对象，键是问题名，值是一道题：noul（是/否，可带true / false判据）、choice（criteria为选项名到说明）、score（criteria为2–10个从低到高的等级）。答案按问题名返回。
默认 "{\n \"billing\": {\n \"type\": \"noul\",\n \"instructions\": \"Is this message about billing?\"\n },\n \"tone\": {\n \"type\": \"choice\",\n \"instructions\": \"What is the tone of this message?\",\n \"criteria\": {\n \"angry\": \"An upset or hostile message\",\n \"calm\": \"A neutral or polite message\",\n \"excited\": \"An enthusiastic or eager message\"\n }\n },\n \"urgency\": {\n \"type\": \"score\",\n \"instructions\": \"How urgent is this message?\",\n \"criteria\": [\n \"Can wait\",\n \"Needs attention this week\",\n \"Needs attention today\"\n ]\n }\n}"