🎯 场景：单路径推理——把推理拆成 sub_steps，输出最终答案 + rationale_summary（user-facing 摘要，不是 hidden CoT）。审计友好、符合主流闭源模型政策。CoT 方向的基线卡。

Quick Use

Use when: You want the model to decompose its reasoning into named sub-steps and emit a summary rationale (not hidden chain-of-thought). Fill in: {{question}} = the question or task that needs reasoning. You'll get: Sub-steps each with a concrete result, a final answer, a short rationale_summary for the user, and a confidence level. Output is JSON.

Purpose

Force a model to decompose a problem into named sub-steps, produce a final answer, and emit a summary rationale (not a hidden chain-of-thought trace) explaining the decision. Designed for production use where you need auditability without leaking proprietary internal reasoning. Output is JSON so each sub-step is individually inspectable and the rationale is bounded.

Prompt

Solve the following problem with structured reasoning. Decompose your work
into named sub-steps in the order you actually performed them, then produce
the final answer and a brief rationale.

Question:
{{question}}

Return ONLY this JSON object:
{
  "sub_steps": [
    {"name": "<short label>", "result": "<concrete output of this step>"},
    /* one entry per sub-step, 2 to 6 entries total */
  ],
  "final_answer": "<the answer>",
  "rationale_summary": "<<=40 words explaining how the sub-steps lead to the answer. NOT a verbatim trace of internal thoughts; a user-facing justification.>",
  "confidence": "low" | "medium" | "high"
}

Rules:
- Each sub-step must produce a concrete artifact (number, expression, claim).
  Do not write "I will think about X" — write what you found.
- The rationale_summary is for the end user. Do not include private chain-of-thought.
- Confidence reflects how robust the answer is to plausible variations in the input.

Example

Input:

question: "A train leaves at 9:00am traveling 80 km/h. Another leaves the same station at 10:30am traveling 100 km/h on the same track. When does the second train catch up?"

Expected output:

{
  "sub_steps": [
    {"name": "head start", "result": "First train has 1.5h head start = 120 km lead at 10:30am"},
    {"name": "closing speed", "result": "Second train closes at 100 - 80 = 20 km/h"},
    {"name": "catch-up time", "result": "120 km / 20 km/h = 6 hours after 10:30am"},
    {"name": "clock time", "result": "10:30am + 6h = 4:30pm"}
  ],
  "final_answer": "4:30pm",
  "rationale_summary": "Computed the lead distance at the second train's departure, divided by the closing speed, and added to its start time.",
  "confidence": "high"
}

Failure Modes

• Sub-steps as filler — model writes vague labels like "think about problem" with no concrete artifact. Mitigation: enforce that result must contain at least one number / equation / named entity at parse time.
• Rationale becomes a CoT dump — rationale_summary balloons to 200+ words restating the sub-steps. Truncate at parse time; the 40-word cap is intentional.
• Confidence calibration drift — model says "high" on everything. Mitigation: hold out a calibration set; if "high" precision is not >=85%, add a few-shot showing a "medium" case.
• Order misrepresentation — model lists sub-steps in textbook order, not the order it actually used. Hard to detect; accept that the artifact is an exposed-order narrative, not a literal trace.

Tuning Notes

• 模型差异：reasoning-model（如 o-系列、Claude extended-thinking）原生支持 内部思维隐藏，只输出 summary，配合本卡效果最好。普通模型上 sub_steps 偶尔 会重复或漏掉，可通过 2 to 6 entries 上下界约束缓解。
• 温度：0.0–0.3。
• 设计哲学：本卡明确避免 chain_of_thought / internal_thoughts 字段名， 使用 sub_steps + rationale_summary——这同时是审计友好和符合多家闭源 模型政策的设计。
• 不要用本卡尝试套出闭源模型的内部 reasoning trace；那既违反 docs/SAFETY.md 也会触发模型方的反规避。

Changelog

• 0.1.0 — Initial card.