Cover-Letter Pre-Submission Rules

Deterministic rules for presubmission_check.py. Adapted from paper-audit/references/PRE_SUBMISSION_RULES.md with cover-letter-specific additions and LaTeX-source-only rules removed.

Banned AI-tone patterns (G4-AI*)

Per-term ladder: the same term appearing 2 times is minor, 3+ times is major. (Before, a term was silent below 3 occurrences, so a vocabulary-diverse draft with each promo word used once slipped through entirely — see AI-DIV below.) Inherits from paper-audit canonical list.

BANNED_TONE_PATTERNS = (
    ("AI1", r"\binnovative\b"),
    ("AI2", r"\bpioneering\b"),
    ("AI3", r"\brevolutionary\b"),
    ("AI4", r"\btransformative\b"),
    ("AI5", r"\bbreakthrough\b"),
    ("AI6", r"\bunprecedented\b"),
    ("AI7", r"\bremarkable\b"),
    ("AI8", r"\bsuperior\b"),
    ("AI9", r"\bsurpass(?:es|ed|ing)?\b"),
    ("AI10", r"\bstate[- ]of[- ]the[- ]art\b"),
    ("AI11", r"\bhighlights? the potential of\b"),
    ("AI12", r"\bpaves? the way\b"),
    ("AI13", r"\bprofound challenges?\b"),
    ("AI14", r"\bat its essence\b"),
)

Cover-letter-specific opener clichés (L2)

Triggers on first non-empty line of the letter body (post-salutation):

LETTER_OPENER_CLICHES = (
    r"^\s*we are (?:pleased|excited|delighted|honored) to (?:submit|share)\b",
    r"^\s*we hereby submit\b",
    r"^\s*please find (?:enclosed|attached)\b",
    r"^\s*it is our (?:great )?pleasure to submit\b",
    r"^\s*enclosed please find\b",
)

Severity: minor. These openings flag the letter as low-effort to editors at top journals.

Cover-letter-specific banned phrases (J1)

Marketing or AI-template language with zero editor signal:

• "novel and innovative"
• "groundbreaking"
• "first-of-its-kind"
• "game-changing"
• "paradigm shift"
• "cutting-edge"
• "of great interest"
• "will be of broad interest"
• "the field is in need of"

Triggers at 1+ occurrence. Severity: minor (major when 3+).

Generic-fit phrasings (J4)

Tier 4 of FORBIDDEN_PHRASES.md: phrasings that signal the author did not read the venue. Triggers at 1+ occurrence, minor / P3, each with a "name the specific X" replacement hint:

• "your journal" / "your prestigious journal" → name the journal
• "fits (well) with/into the scope" → name the scope dimension
• "broad readership" → name the reader profile
• "important contribution to the field" → name the contribution category

Mechanical rules

ID	Severity	Check
G1	minor	Em dash in reader-visible prose (an AI-tone surface signal, aligned with the ISSUE_SCHEMA minor tier).
G2	minor	Paragraph longer than 120 words or 6 sentences (cover letters are shorter than papers — paragraph cap is tighter).
G3	minor	Paragraph starts with weak transition (however, moreover, in addition, furthermore, also).
G4	minor / major	Same banned AI-tone term appears 2 times (minor) or 3+ times (major).
L1	major / minor	Letter exceeds template's word_limit by ≥20% (major) or up to 20% (minor).
L2	minor	First content line matches a known opener cliché.
J1	minor / major	Cover-letter-specific banned phrase appears (minor) or appears 3+ times (major).
J4	minor	Generic-fit phrasing (Tier 4) appears; emits a "name the specific X" replacement hint.
AI-DIV	minor / major	minor when 3 distinct banned AI-tone terms appear (each perhaps once), major at 4+. Catches diverse AI slop the per-term G4 ladder misses.
S1	minor	3 consecutive body paragraphs open with the same first two word tokens (parallel/templated cadence).
S2	minor	Sentence-length coefficient of variation < 0.25 over ≥8 sentences (suspiciously uniform cadence / low burstiness).

Structural AI-trace checks (AI-DIV, S1, S2)

These three checks port the structural AI-trace idea used by latex-paper-en / typst-paper (deai) into the cover-letter genre. They are all minor/P2 except AI-DIV, which escalates to major at 4+ distinct terms, and they report only — no rewriting.

Thresholds are fixed module constants in presubmission_check.py (AI_TONE_DIVERSITY_*, PARALLEL_OPENING_*, SENTENCE_UNIFORMITY_*), not a per-tier YAML: a cover letter is a single short genre, and the scans must run even without a --journal (so no word_limit baseline exists). Length-adaptive thresholds would add a "what baseline when no template?" dimension for no real gain, so we chose deterministic constants that the zero-false-positive fixture (evals/fixtures/human_letter.md) guards.

Parameters (letter-domain tuning vs. the paper-side defaults):

• AI-DIV — counts how many distinct BANNED_TONE_PATTERNS fire at least once. Distinct from G4 (repetition vs. diversity); both may fire on one letter.
• S1 (parallel openings) — window of 3 consecutive paragraphs, opening key = first 2 word tokens (same pins as the paper-side _check_burstiness). The Dear … salutation paragraph is skipped; declaration paragraphs ("We confirm …", "We declare …") differ under a 2-token key, so they do not false-positive.
• S2 (sentence-length uniformity) — whole-letter (cover letters have no sections), gated at ≥8 sentences with CV < 0.25. The paper-side default is min 5 / CV < 0.30 and tier-gated; the letter sample is short and CV-noisy, so the gate is raised and the threshold tightened to suppress false positives.

Not ported from the paper-side deai structural shell:

• throat-clearing paragraph leads — the letter genre already has equivalents (L2 opener clichés + G3 weak-transition paragraph starts); adding it would double-report.
• low-information-density — a cover letter's declaration paragraph is legitimately templated and evidence-marker-free, so this check would false-positive on well-formed letters at ≤400 words.

Required declaration rules (D-<kind>)

Driven by the active template's required_declarations frontmatter list. The check is:

1. Parse the template's required_declarations and optional_declarations arrays.
2. For each required_declarations item with a known detector, scan the letter body for one of the canonical phrasings (see below).
3. If absent, emit a major finding (D-<kind>) with the declaration name.
4. For optional_declarations with a known detector, emit a minor advisory (D-<kind>-opt) if absent.
5. A required kind with no detector emits an informational D-<kind>-unknown (minor/P3, "verify manually") instead of a false "absent" major; an optional kind with no detector is skipped silently.

Canonical phrasings (regex, case-insensitive):

originality:
  - not (?:been )?published elsewhere
  - not under (?:concurrent )?(?:consideration|review|submission)
  - original (?:research|work|manuscript)

dual_submission:
  - not (?:currently )?(?:under (?:concurrent )?(?:consideration|submission|review)|submitted)(?:\s+elsewhere)?
  - single submission policy
  - (?:dual|multiple) submission
  - not (?:been )?submitted (?:to|elsewhere)
  - concurrent consideration

competing_interests:
  - (?:no |declare(?:s)? (?:no |the following )?)?competing interests?
  - conflicts? of interest
  - declare(?:s)? no (?:competing|conflict)

data_availability:
  - data (?:will be |are |is )?(?:made )?available
  - code (?:will be |is )?(?:made )?available
  - materials (?:are|will be) available
  - data and code
  - data availability statement

ethics_irb:
  - institutional review board
  - \bIRB\b
  - \bIACUC\b
  - ethics? (?:committee|approval|board)
  - clinical trial (?:registration|number|identifier)
  - informed consent

authorship:
  - all authors (?:have )?approved
  - all authors (?:have )?read and approved
  - authorship agreement

ai_disclosure:
  - generative ai
  - \bgen[- ]?ai\b
  - (?:used|use of|using|employed|with|disclos\w+|assisted by) (?:a |an |the )?(?:large language model|llms?|generative ai|ai (?:tool|assistant|writing))
  - no (?:generative )?ai (?:tool|assistance|was|were|used)
  - ai[- ](?:assisted|generated) (?:writing|text|content|editing)
  - \b(?:chatgpt|gpt-\d|copilot|gemini)\b

prior_presentation:
  - (?:previously |earlier )?(?:presented|published|appeared) (?:as |in )?(?:a |an )?(?:poster|abstract|preprint|workshop|preliminary|short version)
  - \ba (?:preliminary|prior|earlier|conference) version\b
  - \bpresented at\b
  - \bextends? (?:our|a) (?:prior|earlier|previous) (?:conference |workshop )?(?:paper|version)

ai_disclosure is required by the Nature / Science / Cell templates (ICMJE Jan 2026 Section V; Science requires it in the cover letter explicitly). IEEE / ACM place AI disclosure in the manuscript, so their templates do not list it. Declaration kinds without a detector (excluded_reviewers, artifact_evaluation, reproducibility_statement) are handled by rule 5 above.

Length check (L1)

Reads word_limit from the template. Counts visible words (excludes salutation, address block, signature). Severity ladder:

• 0-100% of limit: OK.
• 100-120% of limit: minor.
• 120%+ of limit: major.

Paragraph shape (G2, G3)

Stricter than paper-audit because cover letters are shorter:

• Long paragraph: >120 words OR >6 sentences (vs. 180/8 for papers).
• Weak topic starter: ≥60 words AND ≥2 sentences AND opens with a weak transition.

Caption / equation / label / citation rules — NOT applicable

Cover letters do not contain LaTeX captions, numbered equations, or label/ref pairs. The corresponding rules from paper-audit are intentionally omitted.