/submission-audit

name: submission-audit
description: Use when a manuscript is close to submission or resubmission and you need a preflight audit for claim support, figure-panel coverage, legend sync, methods references, terminology stability, and venue-facing risks.

Submission Audit

Overview

Use this skill for late-stage manuscript QA. It is narrower than `manuscript-optimizer`: do not use it to redesign a paper from scratch. Use it when the structure mostly exists and the main task is to catch the failures that survive normal revision cycles.

The core rule is simple: never treat a clean-looking manuscript as submission-ready until the front half, figures, legends, methods, supplement, and venue expectations have been checked against each other.

Use the helper script when you want a fast local pass over figure citations:

python ~/.agents/skills/submission-audit/scripts/check_figure_refs.py path/to/manuscript.md
# Claude Code (global install): replace ~/.agents/skills with ~/.claude/skills
# Claude Code (project-local install): replace ~/.agents/skills with .claude/skills

When To Use

Use this skill when:

• The draft is near submission, resubmission, or internal circulation
• Figures and legends are mostly finalized
• The paper needs a last pass for overclaim, missing references, or cross-section drift
• A revision round compressed the prose and may have dropped supporting detail
• The supplement exists and may no longer match the main text

Do not use this skill for:

• Early brainstorming
• Initial section drafting
• Citation discovery from scratch
• Heavy structural rewrites that belong in `manuscript-optimizer`

Audit Order

1. Front-half alignment

• check title, abstract, introduction, and discussion against the actual Results
• flag any claim stronger than the downstream evidence
• check the broad positions against the Readers section of `paper-workflow`'s

`references/section-contracts.md`: a front matter written for specialists is a reader-friction finding, not a polish item 2. Figure and legend coverage

• verify that every main-figure panel and supplementary panel cited in the paper actually exists
• verify that panel letters, metrics, datasets, and numbers agree across figure, legend, and main text

3. Methods and supplement anchoring

• check that methods are cited where needed from Results
• check that supplementary figures, tables, and notes are referenced precisely enough to be usable

4. Terminology and metrics

• enforce one canonical name per concept
• check abbreviations, metric naming, domain-shift labels, cohort names, and model names

5. Risk pass

• overclaim
• evidence gaps
• unsupported mechanism language
• venue-specific style drift
• Statistical analysis declares only statistics the paper reports (see `stats-reporting-audit`)
• voice: audit, defensive or self-critical writing in the main text, SI, legends and data

statements (run the `anti-defensive-writing` detection pass) 6. Venue preflight: what the target venue's editors and reviewers check

• Nature Portfolio:
• reporting-summary readiness
• data and code availability statements
• accession IDs, repositories, and disclosure of sharing restrictions
• image-integrity and raw-data readiness
• end matter: competing interests, author contributions whose initials match the byline, a

generative-AI statement in the Methods scoped to actual use with no AI-generated images, and code availability consistent with the released repository

• preprint, related-manuscript, and conference-proceedings disclosure
• ML conference (`conference-paper-writing`, What Reviewers Check, in the optional set installed

with `--set all`):

• one statement that test data played no part in model or configuration selection
• the number of runs behind each reported number, and their spread
• no configuration chosen on evaluation data without the default configuration's result beside it
• reproducibility and LLM-use statements as the current year's guide asks

Initials are checked against full names, not treated as unique identifiers. A missing or ambiguous contribution mapping is a finding for the authors. It is not authorization to remove a byline entry, change correspondence or assign a contribution role. 7. Source hygiene

• LaTeX comments become public when the source leaves the authors, on arXiv or as supplementary

material

• strip internal paths and notes with `arxiv_latex_cleaner` or an equivalent step, and confirm

the cleaned source compiles 8. Reviewer-side rejection pass

• contribution sufficiency
• writing clarity and reproducibility
• empirical strength
• evaluation completeness
• design or framework soundness

A finding from this pass is fixed with evidence, a precise definition or a narrower claim, never with an added caveat. Anything this audit adds to the manuscript is held to rule zero of `anti-defensive-writing`, because this audit runs after the posture pass.

Required Checks

• Does every substantive abstract claim map to a figure, table, or supplement item?
• Does every Results subsection cite the correct panel range?
• Does every figure legend still reflect the current plot content?
• Are `Methods` cross-references present where interpretation depends on setup or metric definition?
• Is the supplement indexed precisely enough, including panel letters when needed?
• Are strong causal or mechanism words used only where direct evidence exists?
• Are title, abstract, and discussion consistent about the paper's actual contribution type?
• Could a scientist outside the subfield say, from the title and abstract alone, what problem was addressed, what was found and why it matters?
• If the target is `Nature Portfolio`, are the reporting-summary inputs, data/code statements, image-integrity materials, and disclosure items actually ready