stylometric-de-ai-rewriter v1.0 Universal

Stylometric De-AI & Anti-Provenance Rewriter

A self-contained, zero-dependency LLM skill that sanitizes technical AI provenance marks (invisible Unicode, exotic spaces, citation/UTM artifacts, and C2PA/EXIF/XMP/OOXML metadata in Layer A) and rewrites robotic AI prose into natural human writing across all languages and genres using a 6-pass stylometric protocol (Layer B).

1. Core Safeguards

1. Source Fidelity First

Never invents facts, numbers, dates, named sources, citations, or physical details absent from the source text (fiction is exempt).

  • Preserves genuine epistemic hedges (may, might, suggests, is associated with, ehtimal olunur, может) when backed by empirical or clinical uncertainty.
  • Strips only performative throat-clearing ("it is worth noting that").

2. Genre & Register Gate

Classifies the input before editing so essay rules are never forced onto non-essay formats:

  • Emails & replies: Keeps greetings, sign-offs, and actionable task/schedule bullets; leads directly with the decision or question (#26).
  • Academic abstracts: Preserves objective, method, results, conclusion, and confidence bounds.
  • Technical docs & legal text: Keeps numbered steps, lists, tables, code blocks, paths, and exact defined terms.

3. Morphosyntactic Scaling

Adapts sentence-length variation to the morphological typology of the target language:

  • Analytic (EN, ES, FR, PT): Alternates 3–7 word and 35–55 word sentences (CV >= 0.50).
  • Agglutinative & Inflected (AZ, TR, FI, HU, RU, DE, AR): Alternates 3–5 word finite clauses with 22–35 word multi-clause sentences (CV >= 0.45), respecting native SOV/constituent scrambling.
  • Isolating & Logographic (CJK, TH, KM): Varies clause character counts sharply across punctuation boundaries.

2. Tier 1: Layer A — Multi-Vendor Technical Provenance Hygiene

During token regeneration, the skill emits clean canonical characters and strips mechanical provenance markers across OpenAI/ChatGPT, Anthropic/Claude, Google/Gemini, DeepSeek, Qwen, Perplexity, Meta/Llama, Mistral, Cohere, Grok, and Copilot:

Category Target Codepoints & Artifacts Sanitization Action
Unconditional Invisible Strip \u200B (ZWSP), \u2060, \u00AD, \uFEFF (BOM), \u034F, \u061C, Mongolian selectors (\u180B–\u180E), Bidi overrides/isolates (\u200E–\u200F, \u202A–\u202E, \u2066–\u2069), invisible math (\u2061–\u2064), Plane 14 Tags (\uE0000–\uE007F), stego selectors (\uE0100–\uE01EF) Omitted completely during token regeneration.
Conditional Joiner Preservation \u200D (ZWJ) and \u200C (ZWNJ) Preserved in multi-codepoint emoji sequences and orthographically required scripts (Arabic, Persian, Indic, Southeast Asian); stripped when wedged between Latin, Cyrillic, or Greek letters.
Exotic Whitespace & CJK Punctuation Leakage Narrow no-break space (\u202F), \u00A0, \u3000, quad/thin/hair spaces (\u2000–\u200A, \u205F); leaked fullwidth punctuation (，, ：, ；, （, ）) Normalized to standard ASCII space (0x20) and standard ASCII punctuation inside non-CJK sentences.
Multi-Vendor UI & Citation Artifacts 【4:0†source】, :contentReference[oaicite:0]{index=0}, unanchored footnote brackets, ?utm_source=chatgpt.com|claude|gemini|perplexity|copilot Stripped cleanly while preserving valid destination URLs.
Canonical NFC & Mixed-Script Homoglyph Flagging Single word tokens mixing Latin, Cyrillic, and Greek characters (e.g., Cyrillic о inside a Latin word) Normalized to canonical NFC; mixed-script tokens are flagged in the audit table and corrected to pure-script words in the rewrite.

Local Container & File Metadata Sanitization Commands

# PDF (.pdf) — Strip /Info dictionaries, XMP streams, and pdfProducer tags:
exiftool -all:all= input.pdf -o temp.pdf && qpdf --linearize temp.pdf cleaned.pdf

# Office OpenXML & ODF (.docx, .pptx, .xlsx, .odt) — Clear docProps/core.xml & app.xml or run:
exiftool -all= document.docx

# Images & Media (.png, .jpg, .webp, .avif, .mp4) — Strip C2PA JUMBF, EXIF, XMP & PNG chunks:
exiftool -all= image.png

3. Tier 2: Layer B — 6-Pass Stylometric Editing Pipeline

After Layer A hygiene, the draft undergoes a top-down (macro-to-micro) 6-pass transformation addressing all 38 universal LLM signatures. Visit the 6 Core Rules page for architectural details or explore the 13-Language Showcase for live transformations.

Pass Pipeline Stage Signatures Addressed
Pass 1 Macro-Discourse and Epistemic Alignment #4, #21, #22, #25, #27, #28, #29, #30, #31, #32
Pass 2 Typographic and Visual Scaffolding #33, #34, #35, #36, #37, #38
Pass 3 Rhetorical Templates and Transitions #11, #20, #23, #24, #26
Pass 4 Syntactic Tree Variance and Burstiness #2, #3, #7, #8, #9, #10, #12
Pass 5 Lexical Precision and Collocation De-Smoothing #1, #5, #6, #13, #14, #15, #16, #17, #18, #19
Pass 6 Self-Audit Verification Gate Sentence length range, dash quota, punctuation variety, structural tell audit

4. Output Modes

Default Mode

Outputs the final sanitized and rewritten text first, followed by a compact Anti-Provenance & Stylometric Audit table summarizing Layer A artifacts stripped and Pass 6 structural verification metrics (shortest/longest sentence word counts, dash count, and key signatures removed).

Clean / Text-Only Mode

When the user asks for "only the text" or invokes the skill inside an automated pipeline, it outputs strictly the final rewritten text with no commentary or tables.