AI-Run Modern LaTeX Manuscript Workflow and Replication Packet

This record contains a public, reusable workflow note and sanitized replication material for an AI-run, human-directed pipeline that modernizes mathematical manuscripts into source-checkable LaTeX, reader PDFs, translations, audit packages, GitHub mirrors, and Zenodo releases. 2026-06-09 witness-packet refinement. This version adds an SGA/Deligne/Gordan witness-packet addendum. The current scalable pattern is local-to-web: local machines prepare source slices, page maps, OCR prose witnesses, 400-600 dpi formula/diagram/table crops, current TeX anchors, render checks, checksums, and status ledgers; web review threads use those packets for translation, repair, and self-audit against the source. Source-compression rule. A draft may be readable and plausible while silently skipping prose, flattening a commutative diagram, normalizing a symbol, or compressing a displayed argument. OCR prose witnesses are useful as gap detectors, not as authority. Promoted repairs must still be checked against the source scan or high-resolution crop. High-resolution evidence rule. Diagram-heavy and formula-heavy pages often need high-detail evidence. A 150-200 dpi full-page image can provide page context, but cumulative diagram sequences, commutative diagrams, long arrows, small subscripts, accents, table rules, and dense symbolic displays often require labelled 400-600 dpi object crops. Scan-quality rule. Bad scans or overcompressed images can create plausible but wrong outputs. The Gordan repair lane reinforced that OCR and crop witnesses become useful only when the source scan is good enough to support the claim. Before repeating a costly model review pass, check source quality and obtain a better scan or crop when possible. Corrections and improvements can be suggested through GitHub issues or pull requests: https://github.com/KokunoYumeto/modern-latex-manuscripts.