Quick Reference

Read next:

• references/MODE_GUIDE.md — full per-mode workflow, phase steps, committee focus routing
• references/PRESUBMISSION_GUIDE.md — PRESUBMISSION mode-integration matrix

Modes

Mode	Purpose
quick-audit	fast readiness screen with PRESUBMISSION mechanical checks
deep-review	reviewer-style structured critique; Phase 0 includes PRESUBMISSION
gate	PASS/FAIL submission gate; Major/Minor mechanical findings are advisory
re-audit	compare current paper against earlier audit, including mechanical regressions
polish	precheck before a polishing workflow

Legacy aliases:

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

CLI

python audit.py <file> --mode quick-audit
python audit.py <file> --mode deep-review --scholar-eval --literature-search
python audit.py <file> --mode gate --format json
python audit.py <file> --mode re-audit --previous-report old_report.md
python pre_submission_check.py <file> --json

PRESUBMISSION layer

Runs inside quick-audit, gate, re-audit, and deep-review Phase 0.

• Module name: PRESUBMISSION
• Source rule file: references/PRE_SUBMISSION_RULES.md
• Script: scripts/pre_submission_check.py
• Gate behavior: Critical blocks; Major/Minor stay advisory
• Deep-review behavior: full/editor can promote high-signal findings to pre_submission_readiness; focused reviews keep them in Phase 0 context
• PDF behavior: text-only checks; LaTeX/Typst source hygiene is explicitly skipped

Deep-review scripts

python prepare_review_workspace.py paper.tex --output-dir ./review_results
python consolidate_review_findings.py ./review_results/paper-slug
python verify_quotes.py ./review_results/paper-slug --write-back
python render_deep_review_report.py ./review_results/paper-slug
python diff_review_issues.py old_final_issues.json new_final_issues.json

Main outputs

• review_report.md
• revision_suggestions.md
• review_report.html
• revision_suggestions.html
• artifacts/data/final_issues.json
• artifacts/summary/overall_assessment.txt
• artifacts/summary/peer_review_report.md

Common Misclassifications

A short list of recurring LLM errors during paper-audit runs. When in doubt, consult the cited reference and resolve in favor of the more conservative classification.

• Promoting a single em-dash to a gate blocker. PRE_SUBMISSION_RULES G1 treats em-dash overuse as Major only above the threshold. A single em-dash is not a gate-blocking finding. See PRE_SUBMISSION_RULES.md G1.
• Re-finding a section issue inside a cross-cutting lane. If a finding already lives in section_methods or section_results, it should not be re-reported by claims_vs_evidence or notation_and_numeric_consistency unless the cross-cutting view adds a new dimension. See CONSOLIDATION_RULES.md.
• Treating ScholarEval N/A as a Major issue. N/A means the dimension was not exercised (often because --literature-search was not requested), not that the paper failed it. See SCHOLAR_EVAL_GUIDE.md.
• Emitting issue severity inside quick-audit. quick-audit produces a readiness screen, not a reviewer-grade verdict. Severity assignment lives in deep-review synthesis. See MODE_GUIDE.md.
• Silently merging singleton CRITICAL findings. Synthesis must preserve singleton CRITICAL findings unless explicitly downgraded by Arbitration Priority 1 (Evidence Principle). See synthesis_agent.md Forbidden Operations.
• Over-merging across review_lane boundaries. Lane provenance must survive consolidation. Findings from different lanes can be linked via root_cause_key without losing their lane attribution. See CONSOLIDATION_RULES.md and ISSUE_SCHEMA.md.
• Treating PDF mode as a degraded LaTeX mode. PDF runs intentionally skip source-only checks (bibliography hygiene, label resolution, compile warnings). Findings that reference LaTeX macros from a PDF input are TROUBLESHOOTING F5. See TROUBLESHOOTING.md.
• Letting EIC Desk Reject override committee consensus in deep-review. EIC is a 90-second pitch screener; its verdict is binding only in gate mode. See editor_in_chief_agent.md and TROUBLESHOOTING F8.
• Inflating issue lists past lane output limits. Each cross-cutting lane has a max-issues budget in REVIEW_LANE_GUIDE.md. Recurring patterns must collapse into one issue with multiple example locations rather than emit one issue per occurrence.