MathModeling-skills: Skill for Claude Code

Purpose

Make human judgment traceable without multiplying decision files.

Canonical Output

methods/Qx/qx_decisions.jsonl

One JSON object per line. Records are append-only.

Use planning/framing_decisions.jsonl for global/pre-Qx framing decisions.

Required Fields

{
  "schema_version": 1,
  "decision_id": "q1_method_choice",
  "decision_type": "method_choice",
  "status": "DECIDED",
  "decided_by": "human",
  "captured_in_mode": "learning",
  "choice": "M2",
  "rationale": "Human-authored reason tied to evidence.",
  "evidence_refs": ["methods/Q1/probes/risk_probe_summary.json"],
  "decided_at": "ISO-8601",
  "supersedes": null
}

Optional structured fields may include confidence, rejected alternatives, round action, claim scope, assumption labels, or fallback activation.

Workflow

1. Receive the human's answer, the choice-card ID, and evidence paths.
2. Preserve the user's meaning and wording. Normalize only structure, identifiers, and whitespace.
3. Verify:
   • the choice is one of the presented options or explicitly records a user-supplied alternative;
   • evidence paths exist;
   • rationale is non-empty and contains no placeholder;
   • the record does not falsely label AI-authored prose as human-authored.
4. Append one JSON line.
5. If revising a decision, append a new record with supersedes; never overwrite history.
6. Update the compact history in qx_method_card.md only when the decision changes method state.
7. Update the manifest gate/status fields when present.

Decision Types

Typical values:

• framing
• method_choice
• fallback_activation
• result_verdict
• stability_verdict
• assumption_necessity
• claim_scope
• package_signoff
• submission_authorization

Staleness

Mark a decision stale only when its cited evidence materially changed:

• append a decision_stale record naming the old decision and changed evidence;
• ask the human to reconfirm through one choice card;
• do not mark decisions stale because unrelated files or formatting changed.