brycewang-stanford/paper-audit

Paper Audit Skill v4.2

paper-audit is now deep-review-first. Its core job is to behave like a serious reviewer: find technical, methodological, claim-level, and cross-section issues; keep script-backed findings separate from reviewer judgment; and return a structured issue bundle plus a revision roadmap.

Use it for audit and review. Do not use it as the first tool for source editing, sentence rewriting, or build fixing.

What This Skill Produces

• quick-audit: fast submission-readiness screen with script-backed findings
• deep-review: reviewer-style structured issue bundle with major/moderate/minor findings
• gate: PASS/FAIL decision calibrated for submission blockers
• re-audit: compare current issue bundle against a previous audit
• polish: precheck-only handoff into a polishing workflow

The primary product is no longer just a score. For deep-review, the main outputs are:

• final_issues.json
• overall_assessment.txt
• review_report.md
• revision_roadmap.md

Do Not Use

• direct source surgery on .tex / .typ
• compilation debugging as the main task
• free-form literature survey writing
• cosmetic grammar cleanup without an audit goal

Critical Rules

• Never rewrite the paper source unless the user explicitly switches to an editing skill.
• Never fabricate references, baselines, or reviewer evidence.
• Always distinguish [Script] from [LLM] findings.
• Always anchor reviewer findings to a quote, section, or exact textual location.
• Be conservative with OCR noise, formatting quirks, and obvious copy-editing trivia.
• Review like a careful reader: understand the author's intended meaning before flagging an issue.

Mode Selection

| Requested intent | Mode | |---|---| | "check my paper", "quick audit", "submission readiness" | quick-audit | | "review my paper", "simulate peer review", "harsh review", "deep review" | deep-review | | "is this ready to submit", "gate this submission", "blockers only" | gate | | "did I fix these issues", "re-audit", "compare against old review" | re-audit | | "polish the writing, but only if safe" | polish |

Legacy aliases still work for one compatibility cycle:

• self-check -> quick-audit
• review -> deep-review

Committee Focus Routing (deep-review)

For deep-review, use the Academic Pre-Review Committee by default. This is a 5-role review pass:

1. Editor (desk-reject screen)
2. Reviewer 1 (theory contribution)
3. Reviewer 3 (literature dialogue / gap)
4. Reviewer 2 (methodology transparency)
5. Reviewer 4 (logic chain)

If the user requests a single dimension, run only the matching committee role(s).

If --focus ... is provided, it overrides keyword inference:

• --focus full (default)
• --focus editor|theory|literature|methodology|logic

Keyword map (English + Chinese):

• editor: "desk reject", "pre-screen", "editor", "EIC", "主编", "预筛", "初筛"
• theory: "theory", "contribution", "novelty", "theoretical dialogue", "理论", "贡献", "创新性"
• literature: "related work", "literature", "research gap", "citation", "文献", "综述", "Research Gap", "引用"
• methodology: "methods", "sample", "coding", "data", "design", "SRQR", "方法", "样本", "编码", "数据", "研究设计", "透明度"
• logic: "logic", "argument", "causal", "structure", "论证", "因果", "逻辑", "结构"

Output language: match the user's request language. If ambiguous, match the paper language.

Review Standard

Read these references before running reviewer-style work:

1. references/REVIEW_CRITERIA.md
2. references/DEEP_REVIEW_CRITERIA.md
3. references/CHECKLIST.md
4. references/CONSOLIDATION_RULES.md
5. references/ISSUE_SCHEMA.md

The deep-review workflow uses a 16-part issue taxonomy:

1. formula / derivation errors
2. notation inconsistency
3. prose vs formal object mismatch
4. numerical inconsistency
5. missing justification
6. overclaim or claim inaccuracy
7. ambiguity that can mislead a careful reader
8. underspecified methods / missing information
9. internal contradiction
10. self-consistency of standards
11. table structure violations
12. abstract structural incompleteness
13. theory contribution deficiency
14. qualitative methodology opacity
15. pseudo-innovation / straw man
16. paragraph-level argument incoherence

Workflow

Common Step 0

Parse $ARGUMENTS and infer the mode if the user did not provide one. State the inferred mode before running commands if you had to infer it.

quick-audit

1. Run:

   uv run python -B "$SKILL_DIR/scripts/audit.py" <paper> --mode quick-audit ...
   

2. Present a concise report:
   • Submission Blockers first
   • then Quality Improvements
   • then checklist items
   • mark quick-audit findings with [Script] provenance
3. If the user clearly wants reviewer-depth critique after the quick screen, escalate to deep-review.

deep-review

Use this as the default reviewer-style path.

Phase 1: Prepare workspace

Run:

uv run python -B "$SKILL_DIR/scripts/prepare_review_workspace.py" <paper> --output-dir ./review_results

This creates:

• full_text.md
• metadata.json
• section_index.json
• claim_map.json
• paper_summary.md
• sections/*.md
• comments/
• references/ (minimal copies for reviewer agents)
• committee/ (committee reviewer artifacts)

Phase 2: Phase 0 automated audit

Run:

uv run python -B "$SKILL_DIR/scripts/audit.py" <paper> --mode deep-review ...

Treat this as Phase 0 only. It supplies script-backed context and scores, not the final review.

Phase 3: Committee + Review Lanes

Phase 3A: Academic Pre-Review Committee (default)

Decide committee focus:

• If --focus ... is provided, use it.
• Otherwise infer from the user request using the keyword map in "Committee Focus Routing".
• If nothing matches, default to full (all five roles).

Dispatch the committee reviewers (in this exact order) and have them write artifacts into the workspace:

1. agents/committee_editor_agent.md
   • write: committee/editor.md
   • write: comments/committee_editor.json
2. agents/committee_theory_agent.md
   • write: committee/theory.md
   • write: comments/committee_theory.json
3. agents/committee_literature_agent.md
   • write: committee/literature.md
   • write: comments/committee_literature.json
4. agents/committee_methodology_agent.md
   • write: committee/methodology.md
   • write: comments/committee_methodology.json
5. agents/committee_logic_agent.md
   • write: committee/logic.md
   • write: comments/committee_logic.json

If subagents are unavailable, run the committee reviewers inline, but keep the same file outputs.

Then write: committee/consensus.md

• include: overall score (1-10), ordered priorities, and the top 3 issues to fix first
• scoring formula:
  • start at 9.0
  • subtract: 1.5 * (# major) + 0.7 * (# moderate) + 0.2 * (# minor)
  • floor at 1.0
  • if Editor verdict is Desk Reject, cap at 4.0

Note: render_deep_review_report.py automatically embeds committee/*.md into review_report.md when present.

Phase 3B: Section and cross-cutting review lanes (coverage)

Read:

• references/SUBAGENT_TEMPLATES.md
• references/REVIEW_LANE_GUIDE.md

Then dispatch reviewer tasks for:

• section lanes
  • introduction / related work
  • methods
  • results
  • discussion / conclusion
  • appendix, if present
• cross-cutting lanes
  • claims vs evidence
  • notation and numeric consistency
  • evaluation fairness and reproducibility
  • self-standard consistency
  • prior-art and novelty grounding

Each lane writes a JSON array into comments/.

If subagents are unavailable, use the built-in deterministic fallback lane pass in scripts/audit.py so the workflow still writes lane-compatible JSON into comments/ before consolidation.

Phase 4: Consolidation

Run:

uv run python -B "$SKILL_DIR/scripts/consolidate_review_findings.py" <review_dir>
uv run python -B "$SKILL_DIR/scripts/verify_quotes.py" <review_dir> --write-back
uv run python -B "$SKILL_DIR/scripts/render_deep_review_report.py" <review_dir>

Consolidation rules:

• merge exact duplicates
• keep distinct paper-level consequences separate even if they share a root cause
• preserve singleton findings unless clearly false positive
• assign comment_type, severity, confidence, and root_cause_key

Phase 5: Present result

Summarize:

• 1 short paragraph overall assessment
• counts of major / moderate / minor issues
• 3 highest-priority revision items
• path to review_report.md and final_issues.json

gate

1. Run:

   uv run python -B "$SKILL_DIR/scripts/audit.py" <paper> --mode gate ...
   

2. EIC Screening (Phase 0.5): Read agents/editor_in_chief_agent.md and perform the editor-in-chief desk-reject screening on the paper's title, abstract, and introduction. This evaluates pitch quality, venue fit, fatal flaws, and presentation baseline. A desk-reject verdict is a gate blocker.
3. Report PASS/FAIL.
4. Present EIC screening results first (verdict + score + justification).
5. List blockers next.
6. Keep advisory items separate from blockers.
7. For IEEE pseudocode checks, make it explicit which issues are mandatory and which are only IEEE-safe recommendations.

re-audit

1. Requires --previous-report PATH.
2. Run:

   uv run python -B "$SKILL_DIR/scripts/audit.py" <paper> --mode re-audit --previous-report <path> ...
   

3. If both old and new final_issues.json bundles are available, also run:

   uv run python -B "$SKILL_DIR/scripts/diff_review_issues.py" <old_final_issues.json> <new_final_issues.json>
   

4. Present:
   • root-cause-aware status labels: FULLY_ADDRESSED, PARTIALLY_ADDRESSED, NOT_ADDRESSED, NEW
   • use structured prior issue bundles when available, but still accept Markdown previous reports

polish

1. Run the audit precheck:

   uv run python -B "$SKILL_DIR/scripts/audit.py" <paper> --mode polish ...
   

2. If blockers exist, stop and report them.
3. Only proceed into polishing if the precheck is safe.

Output Contract

For deep-review, the final issue schema is:

{
  "title": "short issue title",
  "quote": "exact quote from paper",
  "explanation": "why this matters and what remains problematic",
  "comment_type": "methodology|claim_accuracy|presentation|missing_information",
  "severity": "major|moderate|minor",
  "confidence": "high|medium|low",
  "source_kind": "script|llm",
  "source_section": "methods",
  "related_sections": ["results", "appendix"],
  "root_cause_key": "shared-normalized-key",
  "review_lane": "claims_vs_evidence",
  "gate_blocker": false,
  "quote_verified": true
}

Always prefer:

• exact quotes over vague paraphrase
• evidence-backed findings over style commentary
• issue bundle + roadmap over raw script dumps

References

| File | Purpose | |---|---| | references/REVIEW_CRITERIA.md | top-level audit scoring and mapping | | references/DEEP_REVIEW_CRITERIA.md | deep-review-specific issue taxonomy (16 dimensions) and leniency rules | | references/CONSOLIDATION_RULES.md | deduplication and root-cause merge policy | | references/ISSUE_SCHEMA.md | canonical JSON schema | | references/REVIEW_LANE_GUIDE.md | section lanes and cross-cutting lanes | | references/SUBAGENT_TEMPLATES.md | reviewer task templates | | references/QUICK_REFERENCE.md | CLI and mode cheat sheet |

Scripts

| Script | Purpose | |---|---| | scripts/audit.py | Phase 0 audit and mode entrypoint | | scripts/prepare_review_workspace.py | create deep-review workspace | | scripts/build_claim_map.py | extract headline claims and closure targets | | scripts/consolidate_review_findings.py | deduplicate comment JSONs | | scripts/verify_quotes.py | verify exact quote presence | | scripts/render_deep_review_report.py | render final Markdown report | | scripts/diff_review_issues.py | compare old vs new issue bundles |

Reviewer Lanes

Committee agents (deep-review default):

• committee_editor_agent.md
• committee_theory_agent.md
• committee_literature_agent.md
• committee_methodology_agent.md
• committee_logic_agent.md

Default deep-review lanes live in agents/:

• section_reviewer_agent.md
• claims_evidence_reviewer_agent.md
• notation_consistency_reviewer_agent.md
• evaluation_fairness_reviewer_agent.md
• self_consistency_reviewer_agent.md
• prior_art_reviewer_agent.md
• synthesis_agent.md
• editor_in_chief_agent.md — EIC desk-reject screener (used in gate mode)

Specialized deep-review agents (read their files for activation criteria):

• critical_reviewer_agent.md — devil's advocate with C3-C5 checks
• domain_reviewer_agent.md — domain expertise with A1-A7 assessments
• methodology_reviewer_agent.md — methodology rigor with B3-B10 checks
• literature_reviewer_agent.md — evidence-based literature verification (optional, --literature-search)

Examples

• “Review this manuscript like a serious conference reviewer and tell me the biggest validity risks.”
• “Run a quick audit on paper.tex and tell me what blocks submission.”
• “Gate this IEEE submission and separate blockers from recommendations.”
• “Re-audit this revision against my previous report.”