style-fingerprint-gate

name	style-fingerprint-gate
description	Use before delivering formal academic, reviewer-facing, stakeholder-facing, or submission-facing prose to scan and reduce overused binary negative-contrast constructions such as "rather than", "not...but", "不是...而是", and "而不是"; runs after self-review and before final delivery.

Style Fingerprint Gate

Use this skill for proposals, literature reviews, methodology sections, manuscripts, reports, theses/dissertations, supervisor/PI-facing drafts, reviewer-facing drafts, bilingual drafts, and formal academic prose.

Purpose

Catch repeated binary negative-contrast constructions before delivery. This is a deterministic style-density check, not an AI detector and not a detector-evasion workflow.

Required Position

Run after:

• academic-self-review-loop
• academic register / authorial voice drafting

Run before:

• final artifact generation
• pre-delivery lock
• scripts/formal_delivery_guard.py
• final delivery summary

Workflow

1. Run the scanner on the source Markdown or extracted text:

   python3 scripts/style_fingerprint_scan.py <draft> --strict
   

   The scanner checks three failure modes:

   • density per 1000 estimated words;
   • total number of negative-contrast hits;
   • repeated use of the same pattern category, even in a long document.

2. If density is above threshold:

   • locate each flagged construction;
   • keep only genuine scope distinctions;
   • rewrite mechanical instances into direct positive statements, conditional clauses, or plain definitions;
   • avoid changing argument structure, evidence meaning, citations, or uncertainty level.

3. Re-run the scanner.

4. If any downstream skill edits the prose after the scan, re-run the scanner on the final source text before Word generation or delivery.

5. Record before/after density in the writing checkpoint or delivery checkpoint.

Pattern Inventory

The scanner looks for repeated surface patterns that caused the Production failure and nearby variants:

English:

• rather than: common mechanical contrast marker in the failed draft;
• not just ... but: repeated emphasis-stacking pattern;
• not only ... but: repeated emphasis-stacking pattern;
• it's not ... it's: conversational binary reframing;
• isn't about ... it's about: conversational binary reframing;
• not ... but rather: formal binary reframing;
• not ... but: generic binary reframing, counted only when it does not overlap a more specific match;
• as opposed to: contrast marker often used where a direct scope statement would be cleaner;
• instead of ... it: indirect contrast pattern that often creates a foil before the main claim.

Chinese:

• 并不是...而是
• 并非...而是
• 不是...而是
• 不是...而不是
• 而不是
• 而非
• 不仅...而是
• 与其说...不如说
• 不在于...而在于

The scanner suppresses overlapping matches on the same line so one phrase is not counted twice when a specific pattern already covers it.

Threshold Rationale

Default hard thresholds:

• density: more than 3 hits per 1000 estimated words;
• total hits: more than 12 hits;
• single category repetition: more than 6 hits in one category.

Default review threshold:

• density above 1.5 hits per 1000 estimated words creates a REVIEW status;
• repeated use of one category above the review limit also creates a REVIEW status.

Rationale:

• Repeated binary negative-contrast templates can make research prose feel mechanical even when the evidence is sound.
• Some long documents can hide a repeated template under a tolerable density score, so category caps are also required.
• One or two contrastive phrases can be academically legitimate, especially for scope and limitation.
• Above three hits per 1000 estimated words, the pattern becomes visible enough to require human review before formal delivery.

This threshold is a project style gate, not an institutional, journal, funder, or client requirement.

Scanner Exclusions

The scanner skips:

• fenced code blocks;
• top YAML frontmatter;
• blockquote lines;
• final References, Bibliography, Reference list, Works cited, or 参考文献 sections, whether written as Markdown # headings, plain standalone headings, or setext-style headings.

It may still flag inline quotations or unusual formatting. Treat those as human-review items, not automatic rewrite instructions.

Rewrite Preferences

Prefer:

• direct positive statement: say what the research does or shows;
• conditional boundary: state where a claim applies;
• plain definition: define the concept without first rejecting a foil.

Avoid repeated:

• rather than
• not ... but
• not only ... but
• 不是...而是
• 而不是
• 并不是...而是

Bilingual Rule

For bilingual output, fix the English source first. Then regenerate the Chinese from the cleaned English so the Chinese does not inherit rather than -> 而不是 translation patterns.

Override Path

If a legitimate contrast-heavy passage is blocked, do not silently bypass the gate. Record the reason in the writing or delivery checkpoint and, for formal delivery, use scripts/formal_delivery_guard.py --acknowledge-override --override-reason "<reason>" only after the user explicitly accepts the remaining style risk.

Boundaries

• Do not delete every negation.
• Do not remove legitimate academic scope distinctions.
• Do not change source claims, citations, methods, compliance wording, or verified requirements.
• Do not promise that the text will pass an AI detector.
• If the scan finds prompt residue, placeholders, fake references, or misleading AI-use disclosure, route to academic-integrity-preflight.