Purpose

At every modeling-judgment point, surface the minimal set of questions only the modeler can legitimately answer, before the AI shows any suggestion — and refuse to answer them. This is the reusable form of the one good pattern that already exists in method-selector (the "前置颗粒度对齐 — ask the ONE most load-bearing question" step): make it available at every judgment gate, not just method selection.

The posture this enforces: instead of "AI proposes a verdict → human confirms it" (which collapses into rubber-stamping), the flow becomes "AI asks the question the human owns → human answers → AI assists with what follows". The AI is allowed to lay out the trade-off; it is not allowed to pick the side.

This skill does NOT answer its own questions, pre-select a recommended option, or fill any decision artifact. It produces questions; the human's answers flow into the modeler_decision / modeler_rationale fields of the relevant decision artifact or the decision log.

When to use

Call this skill FIRST, before the judgment-bearing skill shows its evidence/suggestion, at any gate whose verdict is a modeling judgment:

• G1 framing — before problem-classifier shows ai_suggested_type: what output does the team want to defend?
• G2 / G2.5 method choice + baseline — before method-selector shows its ai_suggestion.
• post-experiment (G4.5) — before result-report-generator / robustness-checker show their suggested verdicts: which result is the headline, how confident, is it robust enough?
• G5 claim scope / figure role — before paper-section-writer / figure-table-planner finalize: what is the contribution, what claim does this figure defend?

Any skill whose tier is makes_modeling_judgment (the five C-layer skills) should trigger this skill first. B-layer skills may use it for their load-bearing span (the rubric, the framing, the physical meaning).