Pre-Submission Mechanical Rules

Source attribution: adapted from ref/Supervisor-Skills/plugins/phd-research/skills/pre-submission-reviewer (license: CC-BY-4.0). This reference keeps the rule structure and severity calibration while fitting paper-audit's deep-review-first architecture.

This file lists the deterministic rules and the AI-tone term list. For how those rules plug into each mode (gate behavior, deep-review promotion, PDF skip rules, regression in re-audit), see references/PRESUBMISSION_GUIDE.md.

Purpose

paper-audit uses this layer as a deterministic final-week check, not as a competing mode. The PRESUBMISSION script feeds:

• quick-audit: fast mechanical readiness signals.
• gate: advisory findings plus any true Critical blockers.
• re-audit: regression comparison for mechanical findings.
• deep-review Phase 0: context for reviewer lanes.

For full/editor deep-review, high-signal PRESUBMISSION findings can become pre_submission_readiness lane issues. Focused theory, literature, methodology, and logic reviews keep these findings in Phase 0 context only.

Severity Mapping

Source taxonomy	paper-audit script severity	deep-review bundle severity
CRITICAL	Critical / P0	major, gate_blocker=true
MAJOR	Major / P1	moderate
MINOR	Minor / P2	minor or Phase 0 only

gate fails only on Critical script findings or failed checklist items. Major and Minor PRESUBMISSION findings stay advisory.

Deterministic Script Rules

These rules are safe for pre_submission_check.py because they are directly visible in source or extracted PDF text.

Grammar and style IDs

• G1: em dash found in reader-visible prose. Default severity: Major.
• G2: paragraph longer than 180 words or more than 8 sentences. Default: Minor.
• G3: paragraph starts with a weak transition instead of a claim-like topic sentence. Default: Minor.
• G4: banned AI-tone term group appears three or more times. Default: Major.
• G5: abstract lacks the five submission elements: background, objective, method, results, conclusion. Missing results or missing quantitative result evidence is Major; other missing elements are Minor.
• G6: repeated promotional claim language without local evidence should be escalated to reviewer judgment, not auto-failed.
• G7: article, tense, agreement, which/that, and Chinglish findings require exact quoted evidence when reported by an LLM reviewer.
• G8: passive voice or sentence-length findings are Minor unless they hide a method or result claim.

LaTeX/source hygiene IDs

• L1: citation lacks a non-breaking tie before \cite..., for example Method \cite{key} instead of Method~\cite{key}. Default: Minor.
• L2: label contains spaces. Default: Major.
• L3: label uses hyphens where underscores are safer. Default: Minor.
• L4: numbered equation environment has no label. Default: Minor.
• L5: numbered equation label is never referenced in text. Default: Minor.
• L6: figure/table source has a caption but no concrete finding. Default: Minor.
• L7: figure/table source has a label but no preceding in-text reference. Covered by references; do not duplicate unless source context matters.
• L8: bibliography, citation key existence, and malformed references stay in bib, citations, and references modules.
• L9: page-limit checks stay venue/checklist specific.
• L10: compile errors are outside paper-audit; route to the LaTeX/Typst writing skill when the user asks to fix source.
• L11: source comments should be ignored by mechanical scans.
• L12: PDF input must skip source-only checks instead of guessing LaTeX state.
• L13: source-only findings should include line numbers whenever available.

Figure quality IDs

The script can only check source captions. The following remain manual or LLM reviewer prompts:

• F1: caption lacks a concrete finding or comparison cue. Script-checkable from source captions.
• F2: raster instead of vector final figure.
• F3: post-scaling font size too small.
• F4: non-color-blind-safe palette without dual encoding.
• F5: chartjunk or decorative elements that obscure comparison.
• F6: overview figure labels do not match section titles.

F2-F6 require the PDF/figure artifacts or reviewer inspection. Do not report them as deterministic script findings unless a visual checker proves them.

Forbidden AI-Tone Terms

The script treats these as a frequency threshold, not a single-word ban. Three or more matches in reader-visible text becomes Major/P1:

• innovative
• pioneering
• revolutionary
• transformative
• breakthrough
• unprecedented
• remarkable
• superior
• surpass
• state-of-the-art
• highlights the potential of
• pave the way
• profound challenge
• at its essence

Single uses may be legitimate technical wording. Keep them in Phase 0 context unless they become a pattern.

Abstract Five-Element Rule

A submission-ready abstract should show:

1. Background or problem pressure.
2. Objective or research question.
3. Method or approach.
4. Concrete result, preferably quantitative.
5. Conclusion, implication, or bounded takeaway.

Missing quantitative result evidence is a Major readiness finding because it prevents a final-week gate from seeing what the paper actually achieved.

PDF vs Source Behavior

PDF mode runs only text-verifiable checks:

• em dash scan
• banned AI-tone frequency
• abstract five-element check
• long paragraph/topic-sentence weak signals

PDF mode explicitly skips:

• LaTeX citation tie hygiene
• label space/hyphen rules
• numbered equation reference rules
• source-caption rules

The skip must be visible in script output as an ignored comment or metadata, not as an issue.

Integrity Gate

Before using these findings in reviewer-facing output:

1. Every finding must be tied to source text, a line number, or a section.
2. Severity must follow the mapping above; taste-only concerns are not Critical.
3. Script findings must keep [Script] provenance.
4. PRESUBMISSION findings should not replace deep reviewer judgment about methodology, theory, literature, or claim validity.
5. The skill audits only; it does not rewrite paper source.