decision-prompt-builder

Purpose

Ask the smallest useful question that only the human modeler can answer. Present mutually exclusive options with consequences; do not turn mechanical checks into user questions.

Inputs

• Current gate and the judgment it needs.
• Problem goal, required output, hard constraints, and available evidence.
• planning/session_config.json.
• Existing decisions in methods/Qx/qx_decisions.jsonl.

Configuration

• Read interaction_mode; accept legacy mode for compatibility.
• learning: show 2–3 short questions and withhold the AI suggestion until the user answers.
• speed: show one compressed question and optionally show the AI suggestion alongside.
• rigor_profile does not change who owns the judgment.

Choice-Card Workflow

1. Identify one load-bearing judgment.
2. Create 2–3 mutually exclusive options. Each option must state its practical consequence.
3. Add 都不合适 / 补充约束 when the listed options may not cover the user's intent.
4. Ask no more than three questions in one card.
5. Do not recommend an option in learning mode before the answer.
6. Pass the answer verbatim to modeler-decision-logger; do not create a per-skill pending decision file.

Standard Cards

Before method screening

Ask only the missing high-impact items:

• output form to defend;
• interpretability/performance priority;
• unacceptable failure;
• experiment budget.

Do not ask the user to choose an algorithm name before evidence exists.

Example:

请选择这轮方案的首要取向：

- A. 可解释性优先——方法更透明，但可能牺牲部分拟合效果。
- B. 平衡——接受中等复杂度，要求能解释且优于可信 baseline。
- C. 性能优先——允许更复杂的方法，但需要额外稳健性和解释工作。
- D. 都不合适 / 我补充约束。

After a meaningful experiment

Use computed evidence to ask:

• proceed with the current main method;
• adjust a stated assumption or parameter and rerun;
• activate the recorded fallback.

Name the consequence and evidence for each option. Do not silently convert an AI metric preference into the human verdict.

Before final freeze

Use only when claim scope or confidence is genuinely judgment-bearing:

• keep the claim;
• downgrade it;
• drop it.

Output

Return one choice_card block containing:

• decision_id
• decision_type
• question
• 2–3 options plus optional constraint override
• evidence paths
• the consequence of each option

Do not save the card unless another skill needs a durable prompt record.

Rules

• Ask about trade-offs, not mechanically determinable facts.
• Prefer one card at a decision point; avoid repeated micro-confirmations.
• Do not pre-fill the user's choice or rationale.
• Do not mark a decision DECIDED.
• Do not require a prose essay. One evidence-linked sentence is sufficient when it captures the user's real reason.
• If there is no genuine human judgment, return control without asking a question.

Verification

• Options are mutually exclusive and consequences are clear.
• The card is grounded in the current problem or computed evidence.
• No hidden recommendation appears in learning mode.
• No per-skill decision artifact was created.