Quality Assurance Auditor

Purpose

Audit the complete mathematical modeling contest solution before final submission.

This skill is the third of three independent auditors that together form Gate G6. It runs after consistency-auditor and completeness-auditor and consumes their reports as inputs — it does NOT re-do their checks. Each auditor catches a different failure mode:

• consistency-auditor → cross-media numbers / files / symbols / parameters (does the paper claim match the code?)
• completeness-auditor → audit/review files exist on disk with ≥ 5 pass items (did the producer skills actually write what they claimed?)
• quality-assurance-auditor (this skill) → workflow completeness, three critical rules, anti-fabrication, paper quality (is the whole pipeline coherent and defensible?)

This skill checks whether the solution answers every subquestion, whether models / code / results / figures / robustness checks / paper sections are mutually coherent, whether the full workflow completeness requirements are met, and whether unsupported claims or missing artifacts block final delivery.

This skill verifies that every subquestion has gone through the full delivery chain: candidate methods (with PoC) → experiments → final method explanation → final result analysis → solution package (with frozen_numbers.json) → paper sections.

This skill does not invent missing results, rewrite the full paper, rerun models, or approve submission when evidence is incomplete. It also does not approve assembly if the other two audits (consistency / completeness) are missing or failed — the three are orthogonal and all must converge.

When to use

Use this skill:

• After paper-section-writer has drafted paper sections for all subquestions.
• Before final assembly or submission.
• When checking whether the solution is logically complete.
• When identifying blocking issues, minor issues, and repair actions.
• When deciding whether to return to an upstream skill.

Preconditions

The following should already exist or be provided:

• Parsed problem and subquestions.
• Problem classification.
• Candidate method pools per subquestion.
• Data report and cleaned data summary, if data is used.
• Reviewed model code.
• Model result artifacts per subquestion.
• Robustness or sensitivity reports per subquestion.
• Figure-table plans per subquestion.
• Paper section drafts for all subquestions.

If paper sections are missing, hand back to paper-section-writer.

If model results are missing, hand back to python-model-code-generator, matlab-model-code-generator, or code-reviewer.

Inputs

Use or request:

• workspace/problem/problem-parser/problem_parse.json
• workspace/problem/problem-classifier/problem_classification.json
• methods/Qx/qx_method_candidates.md (all subquestions)
• methods/Qx/qx_method_iteration_log.md (all subquestions)
• methods/Qx/qx_final_method_explanation.md (all subquestions)
• results/Qx/reports/qx_final_result_analysis.md (all subquestions)
• results/Qx/reports/qx_solution_package_for_writer.md (all subquestions)
• Result files under results/Qx/experiments/
• Robustness reports under robustness/Qx/
• Figure-table plans under methods/Qx/
• Figures under results/Qx/experiments/*/figures/ and paper/figures/
• Paper drafts under paper/sections/
• Generated code under code/
• planning/progress_dashboard.md (workflow status)
• Contest requirements or submission rules if available

Workflow

1. Verify sibling auditors have run (independent audit layer prerequisite).

• Check that paper/audits/cross_media_consistency_audit.md exists with verdict PASSED.
• Check that paper/audits/completeness_audit.md exists with verdict PASSED.
• If either is missing, route to it first — do not pretend QA can substitute for them.
• If either has FAILED verdict, treat that as a BLOCKING issue and inherit its repair routing.
• Only when both sibling audits PASSED, proceed to step 1.

QA is the third independent auditor, not a super-auditor that subsumes the others.

1. Check problem coverage.

• Verify every original subquestion is represented in the paper.
• Verify each subquestion has a corresponding model, result, and written answer.
• Mark unanswered or partially answered subquestions as blocking issues.

1. Check workflow completeness (NEW — expanded check).

• For each subquestion, verify the full delivery chain:

1. Candidate method pool exists (methods/Qx/qx_method_candidates.md).
2. Experiment reports exist (at least one round at results/Qx/experiments/roundN/).
3. Method iteration log exists (methods/Qx/qx_method_iteration_log.md).
4. Final method explanation exists (methods/Qx/qx_final_method_explanation.md).
5. Final result analysis exists (results/Qx/reports/qx_final_result_analysis.md).
6. Solution package exists (results/Qx/reports/qx_solution_package_for_writer.md).
7. Robustness report exists (robustness/Qx/qx_robustness_report.md) or a justified exception.
8. Figure-table plan exists (methods/Qx/qx_figure_table_plan.md).

• Mark any missing link in this chain as a blocking issue.

1. Check the three critical rules.

• Rule 1: Verify paper model description matches the final method explanation (not an early candidate pool).
• Rule 2: Verify paper results come from the final result analysis (not raw experiment outputs directly).
• Rule 3: Verify the solution package exists and the paper writer used it.

1. Check artifact lineage.

• Trace each major paper claim back to problem parse, method explanation, data, code, result, figure, or robustness artifact.
• Mark claims without artifacts as unsupported.

1. Check model consistency.

• Verify models match the classified problem types.
• Verify baseline models exist for main models unless explicitly justified.
• Verify the paper describes the final method (from qx_final_method_explanation.md), not an earlier candidate.
• Verify eliminated methods are mentioned appropriately.

1. Check data and result consistency.

• Verify raw data was not overwritten.
• Verify cleaned data and field mappings are documented.
• Verify result tables come from reviewed scripts or accepted computation.
• Verify units, variables, and dimensions are consistent.

1. Check paper quality.

• Verify every referenced figure or table exists on disk.
• Verify figures in the paper match the figure-table plan's Type 3 (论文图) classification.
• Verify no Type 1 (诊断图) figures accidentally appear in the paper.
• Verify every figure or table has a source artifact.
• Verify captions support the intended claim.
• Verify assumptions are necessary and linked to modeling needs.
• Verify notation is defined before use and consistent with the global symbol table.
• Verify conclusions map back to subquestions.
• Verify limitations are not hidden.
• Reject decorative or unsupported visuals.

1. Check for fabrication and exaggeration.

• Verify no references are fabricated.
• Verify no numerical values are invented.
• Verify no experiments are claimed that were not run.
• Verify robustness claims do not exceed completed checks.
• Verify conclusions do not exceed evidence.
• Flag any claim that cannot be traced to an artifact.

1. Produce a QA report — MANDATORY substantive file on disk.

• Save to paper/qa_report.md.
• List ≥ 5 explicit pass items with file:line evidence (concrete checks, not "all good"). These are checked by completeness-auditor.
• Separate blocking issues from minor issues.
• Provide a repair plan.
• Route each repair to the appropriate upstream skill.
• Approve final assembly only when ALL THREE auditors PASS (this skill + consistency-auditor + completeness-auditor). Document the cross-auditor verdict explicitly in the report.
• If sibling auditors are missing or failed (step 0), QA cannot approve assembly — note this and route appropriately.

Outputs

Produce a QA report at:

• paper/qa_report.md

Containing:

• qa_status (passed / failed)
• final_assembly_allowed (true / false)
• blocking_issues
• minor_issues
• workflow_completeness_check (per subquestion)
• three_critical_rules_check
• unsupported_claims
• fabrication_risks
• artifact_traceability
• subquestion_coverage
• repair_plan
• recommended_next_skill

Output format

# QA Report

## Overall Status

- **QA Status**: passed / failed
- **Final Assembly Allowed**: true / false (only if ALL THREE auditors PASS — see below)
- **Date**: [timestamp]

## Sibling Auditor Status (Gate G6 prerequisite)

| Auditor | Report Path | Verdict |
|---------|-------------|---------|
| consistency-auditor | paper/audits/cross_media_consistency_audit.md | PASSED / FAILED / MISSING |
| completeness-auditor | paper/audits/completeness_audit.md | PASSED / FAILED / MISSING |
| quality-assurance-auditor (this report) | paper/qa_report.md | PASSED / FAILED |

**Cross-auditor verdict**: ALL_PASSED (assembly allowed) / ONE_OR_MORE_FAILED (assembly blocked)

## Pass Items (≥ 5 REQUIRED for this audit to count as substantive)

1. ✅ Every subquestion (Q1, Q2, Q3, Q4) has paper/sections/qx.tex on disk and non-empty.
2. ✅ All `frozen_numbers.json` files exist and are newer than their referenced code (no stale freeze).
3. ✅ Every main model has a baseline comparison in its results section.
4. ✅ No fabricated references detected (cross-checked against reference-manager's reference_audit.md).
5. ✅ Every figure in `paper/figures/` is Type 3 (论文图) or Type 4 (附录图); no Type 1 (诊断图) leaked in.
6. ✅ ...

## Workflow Completeness Check

| Check | Q1 | Q2 | Q3 | Q4 |
|-------|----|----|----|-----|
| 1. Candidate method pool | ✅ | ✅ | ✅ | ❌ |
| 2. Experiment report(s) | ✅ (R2) | ✅ (R1) | ❌ | ❌ |
| 3. Method iteration log | ✅ | ❌ | ❌ | ❌ |
| 4. Final method explanation | ✅ | ❌ | ❌ | ❌ |
| 5. Final result analysis | ✅ | ❌ | ❌ | ❌ |
| 6. Solution package | ✅ | ❌ | ❌ | ❌ |
| 7. Robustness report | ✅ | ❌ | ❌ | ❌ |
| 8. Figure-table plan | ✅ | ❌ | ❌ | ❌ |

## Three Critical Rules Check

| Rule | Q1 | Q2 | Q3 | Q4 | Status |
|------|----|----|----|-----|--------|
| Rule 1: Paper method matches final method explanation | ✅ | ❌ (no final explanation) | N/A | N/A | Q2 FAIL |
| Rule 2: Paper results from final result analysis | ✅ | N/A | N/A | N/A | PASS |
| Rule 3: Solution package exists and was used | ✅ | ❌ (no package) | N/A | N/A | Q2 FAIL |

## Subquestion Coverage

| Subquestion | Answered in Paper | Model Exists | Results Exist | Robustness Exists | Status |
|-------------|-------------------|-------------|---------------|-------------------|--------|
| Q1 | ✅ | ✅ (entropy-TOPSIS) | ✅ | ✅ | Complete |
| Q2 | ⏳ (partial draft) | ❌ (no final) | ❌ | ❌ | Blocked |
| Q3 | ❌ | ❌ | ❌ | ❌ | Blocked |
| Q4 | ❌ | ❌ | ❌ | ❌ | Blocked |

## Blocking Issues

| # | Issue | Affects | Why It Matters | Repair Skill |
|---|-------|---------|---------------|--------------|
| 1 | Q2 has no final method explanation | Q2 completion | Paper cannot describe the final model | `final-method-explainer` |
| 2 | Q3 not started | Q3 completion | Subquestion unanswered | `workflow-orchestrator` |
| 3 | Q4 not started | Q4 completion | Subquestion unanswered | `workflow-orchestrator` |

## Minor Issues

| # | Issue | Affects | Repair Skill |
|---|-------|---------|--------------|
| 1 | Q1 Figure 1.3 caption does not state the main takeaway | Q1 paper section | `paper-section-writer` |
| 2 | Q1 robustness report missing boundary condition statement | Q1 robustness section | `robustness-checker` |

## Unsupported Claims

| # | Claim | Location | Why Unsupported | Action |
|---|-------|----------|----------------|--------|
| 1 | "The model is globally stable." | paper/sections/q1.tex | Robustness only checked ±10% weight perturbation | Downgrade to "stable under moderate perturbation" |

## Fabrication Risks

| # | Risk | Location | Evidence |
|---|------|----------|----------|
| — | (none detected) | — | — |

## Artifact Traceability (spot check)

| Paper Claim | Source Artifact | Traceable? |
|-------------|----------------|------------|
| "City A ranks highest with score 0.88." | `results/Q1/experiments/final/tables/m2_scores.csv` | ✅ |

## Repair Plan

1. **BLOCKING**: Run `final-method-explainer` for Q2 to produce `methods/Q2/q2_final_method_explanation.md`.
2. **BLOCKING**: Run `workflow-orchestrator` to plan Q3 and Q4 execution.
3. **MINOR**: Update Q1 Figure 1.3 caption via `paper-section-writer`.
4. **MINOR**: Add boundary condition statement to Q1 robustness report via `robustness-checker`.

## Recommended Next Skill

- **If QA failed**: `workflow-orchestrator` (with repair plan)
- **If QA passed**: `workflow-orchestrator` (for final assembly)

Expanded QA checklist

Workflow completeness (per subquestion)

• [ ] Candidate method pool exists
• [ ] Experiment report(s) exist
• [ ] Method iteration log exists
• [ ] Final method explanation exists
• [ ] Final result analysis exists
• [ ] Solution package exists
• [ ] Robustness report exists (or justified exception)
• [ ] Figure-table plan exists

Three critical rules

• [ ] Paper model description matches final method explanation (not earlier candidate pool)
• [ ] Paper results come from final result analysis (not raw experiment outputs)
• [ ] Solution package exists and was used as primary source

Problem coverage

• [ ] Every subquestion is answered in the paper
• [ ] Every answer has model-result-paper alignment

Model consistency

• [ ] Every main model has a baseline or justified exception
• [ ] Paper describes the final method, not an early draft
• [ ] Eliminated methods are mentioned where appropriate

Evidence quality

• [ ] Every numerical claim has a result artifact
• [ ] Every figure or table has a source artifact
• [ ] Every figure in the paper exists on disk
• [ ] No Type 1 (诊断图) figures in the paper
• [ ] Every robustness claim has robustness evidence

Paper quality

• [ ] Every conclusion maps to a subquestion
• [ ] Assumptions are necessary and stated clearly
• [ ] Notation is defined and consistent (cross-check with global symbol table)
• [ ] Code outputs match paper claims
• [ ] Raw data is preserved
• [ ] Limitations are stated

Anti-fabrication

• [ ] No fabricated data, references, results, figures, or experiments
• [ ] No invented numerical values
• [ ] No unsupported superiority claims
• [ ] No hidden uncertainty

Rules

• Do not fabricate missing evidence.
• Do not approve final assembly with blocking issues.
• Do not approve final assembly unless ALL THREE auditors (consistency / completeness / this skill) PASS. A single auditor's "✅" is not sufficient — the three are orthogonal.
• Do not re-do the work of consistency-auditor (cross-media number/file/symbol checks) or completeness-auditor (audit-file existence checks). Consume their reports as inputs.
• Do not approve "完成" verbally. Write the substantive paper/qa_report.md with ≥ 5 explicit pass items. completeness-auditor will flag a verbal-only QA as MISSING.
• Do not hide uncertainty.
• Do not treat planned artifacts as completed artifacts.
• Do not rewrite the full paper inside QA.
• Do not rerun models unless explicitly asked by the user.
• Do not change upstream artifacts silently.
• Route repairs to the correct upstream skill.
• Keep QA findings specific and actionable.
• Check workflow completeness — it's not just about paper quality.

Verification

Before approving final assembly, verify:

• blocking_issues is empty.
• workflow_completeness_check passes for all subquestions (8/8 checks per subquestion, or justified exceptions).
• three_critical_rules_check passes for all subquestions.
• unsupported_claims is empty or explicitly removed from final text.
• Every subquestion is covered.
• Baseline and robustness requirements are satisfied.

…

Source & license

This open-source skill is cataloged on AgentStack and links to its original source — we do not rehost the code.

• Author: zhnnky329
• Source: zhnnky329/MathModeling-skills
• License: MIT

Install and usage instructions live in the source repository linked above.