<< All versions

Skill v1.0.0

currentLLM-judged scan90/100
zhnnky329/mathmodeling-skills/modeler-decision-logger
──Details
PublishedOctober 22, 2026 at 05:37 AM
Content Hashsha256:70470d8a304ee950...
Git SHA046a6e74814c
──Files
Files (1 file, 3.1 KB)
SKILL.md3.1 KBactive
SKILL.md · 98 lines · 3.1 KB

version: "1.0.0" name: modeler-decision-logger description: Faithfully append a human modeler's choice and rationale to one canonical per-subquestion JSONL decision ledger. Use after a choice card is answered or when migrating legacy decision artifacts; never originate or improve the decision.


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

json
{
"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.
  1. Append one JSON line.
  2. If revising a decision, append a new record with supersedes; never overwrite history.
  3. Update the compact history in qx_method_card.md only when the decision changes method state.
  4. 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.

Legacy Migration

Read legacy:

  • methods/Qx/qx_decision_log.md
  • methods/Qx/decisions/*_modeler_decision.md

Migrate only completed human decisions. Preserve original timestamps and source paths when available. Do not convert PENDING placeholders into decisions.

Rules

  • Never choose, rationalize, strengthen, or complete the user's decision.
  • Never create a separate pending decision artifact.
  • Never copy ai_suggestion into the human rationale.
  • One evidence-linked sentence is enough; do not impose arbitrary prose length.
  • Preserve an honest AI-use provenance distinction.

Verification

  • JSONL is valid one-object-per-line.
  • Record is append-only and uniquely identified.
  • Human ownership and evidence are accurate.
  • Supersession and staleness preserve history.
All versions