Zero AI Signature Pipeline: Detection, Removal, and Quality Gating for Bahasa Indonesia and English Nonfiction

Abstract This paper describes a production pipeline for detecting and removing AI writing signatures from manuscripts before publication. The pipeline targets Bahasa Indonesia and English nonfiction prose. It includes a curated pattern library (30+ signatures), a regex-based detector, an auto-fixer that preserves document formatting, an iterative detect-fix-redetect loop, a voice consistency checker, and a pre-publication quality gate. The pipeline has been used in production across 558+ published titles to achieve a target AI signature score of zero at distribution time. The contribution is not novel detection theory but the discipline of treating zero-signature as a hard publication gate rather than a soft preference, and the explicit construction of a Bahasa Indonesia AI signature library, which is rare in the literature. What this deposit contains paper.md — the methodology paper (the citable scholarly artifact) supplementary.zip — bundled working source code for 8 component workflows covered by this paper, each with its own README, Python source, configuration, sample input, CITATION.cff, and LICENSE README.md — entry point CITATION.cff — machine-readable citation LICENSE (MIT for code) and LICENSE-DOCS (CC BY 4.0 for documentation) Production Results Across 558+ published titles: 100% of distributed titles pass the AI signature gate (score = 0) at compile time. Average detect-fix iterations to reach zero: 2.4. Most common manual-review patterns: rhetorical question clusters, staccato triads, ‘Bayangkan’ openers in originally hand-written drafts that absorbed AI rewrites. Zero customer reports of ‘this reads like AI’ across distribution channels with public review surfaces. Limitations The pattern library is finite and grows. New AI tells appear with new model releases. The library is a living document, not a finished one. The detector is regex-based. Patterns with semantic structure (e.g. ‘rhetorical questions clustered to evade individual matching’) are partially addressed by cluster regex but ultimately need human review. Bahasa Indonesia coverage is stronger than English coverage simply because this author writes more in BI; English patterns reflect a smaller corpus. Reproducibility All eight components are released as MIT-licensed source with the pattern library. The pattern library itself is a JSON file (CC BY 4.0) and may be forked, extended, or merged into other quality-control pipelines. Companion Papers This deposit is one of ten flagship records in the Practitioner Publishing Stack series. Each flagship is a methodology paper plus the relevant working source as supplementary material. The series: I. Markdown to multi-format compile pipeline (.docx, .pdf, .epub) with consistent typography. (slug: f01-pub-compile-stack, domain: PUB) II. (this paper) — Detection and removal of AI writing signatures from manuscripts before publication. III. End-to-end pipeline from literature search to Zenodo-published, DOI-indexed academic artifact. (slug: f03-res-research-to-doi, domain: RES) IV. Six python-docx utilities for format normalization, heading conversion, and AI-signature cleanup. (slug: f04-doc-production-pipeline, domain: DOC) V. Visual design system: color tokens, typography, charts, diagrams, and AI cover composition. (slug: f05-vis-leather-and-steel-design-system, domain: VIS) VI. Markdown-driven template engine producing branded spreadsheets and PDF checklists for direct sale. (slug: f06-tpl-spec-to-sellable, domain: TPL) VII. Per-channel specification workflow for distribution across seven publishing platforms. (slug: f07-dst-multi-platform-distribution, domain: DST) VIII. Course-materials pipeline: markdown syllabus to branded PDF and per-session DOCX. (slug: f08-edu-course-materials, domain: EDU) IX. Seven-phase book production methodology with hard quality gates and 18-day average cycle time. (slug: f09-meta-book-lifecycle, domain: META) X. Architectural overview of an independent publishing operation that produced 558+ titles. (slug: f10-meta-practitioner-publishing-stack, domain: META) Author Ibrahim Anwar (Hibranwar) ORCID: 0009-0006-0425-4923 Wikidata: Q138856145 Web: hibranwar.com Affiliation: PT Hibrkraft Kreasi Indonesia (Cileungsi, Bogor, Indonesia) License The methodology paper, configuration, and sample data are released under CC BY 4.0. The Python source code in supplementary.zip is released under the MIT License. Citation If you use this work, please cite via the DOI minted on this Zenodo record. A machine-readable CITATION.cff ships in the deposit. About the Practitioner Publishing Stack The Practitioner Publishing Stack documents an independent publishing operation by Ibrahim Anwar that produced 558+ titles across nonfiction books, public-domain translations, academic papers, and digital templates, distributed across seven platforms, with a single human as the bottleneck. Average end-to-end cycle time per title: 18 days. Operator headcount: 1.