Purpose

Independent cross-media consistency auditor. This skill exists because the rest of the pipeline cannot be trusted to police itself: every generative skill (method-selector, code-generator, result-report-generator, paper-section-writer, ...) self-declares "完成", and history shows that "完成" often hides drift — a number written into the paper is one bug-fix behind the code, a figure filename in the tex no longer exists on disk, a symbol in the method explanation has been renamed in the symbol table.

This skill compares the same fact stated in multiple places and reports every divergence. It does not believe any single source — it cross-checks them.

The audit covers five dimensions:

1. Numbers — every numerical claim in paper/sections/*.tex is traced back to a result artifact (results/Qx/reports/*.md, results/Qx/experiments/*/metrics/*.json|csv, frozen_numbers.json).
2. File names — every figure/table/appendix code reference in the paper must point to a file that exists on disk.
3. Symbols — symbols used in paper/sections/*.tex and methods/Qx/qx_final_method_explanation.md must be defined in planning/symbol_table.md, with the same meaning everywhere.
4. Parameters — parameter values written into the paper (e.g. K=30, α=0.05, seed=2026) must match the values actually used in code/Qx/*.py or code/matlab/Qx/*.m.
5. Decision provenance — every judgment claim in the paper ("we chose X because…", confidence, verdict) must trace to a DECIDED record in methods/Qx/qx_decision_log.md via a decision_id marker. This is the decision-side analogue of the numbers check; it stops the AI from re-authoring the human's reasoning.
6. Requirement and estimator alignment — hard data-use restrictions, requested outputs, estimator/loss functions, refitting rules, and test definitions must agree across problem parse, method explanation, code, results, and paper.

This skill does NOT write the paper, fix the code, or regenerate results. It only reports divergences and routes them to the right repair skill.

This skill runs in parallel with quality-assurance-auditor and completeness-auditor. The three form an independent audit layer; any one of them failing blocks final assembly.

When to use