PDFMDTools PDF to DOCX reproducibility kit
Run date: 2026-10-13. Inputs are self-made synthetic documents.
Settings: Simple line order (comparison), then Layout-aware reconstruction.
All PDFs: one US Letter page, 612 x 792 points.
Scan: locally generated pixel image only. Both modes reject page 1; no output.
Columns: Helvetica 11pt; x=50/330pt, four numbered lines per column.
Table: Helvetica 11pt; x=58/238/418pt, one header and six numbered rows.
Vertical: Noto Sans SC 18pt, individually positioned CJK glyphs; x=430/370pt, 27pt vertical step. Read right stack then left stack. Output is horizontal.
Before/after means the two settings of the same converter version, not historical deployments.
Actual browser downloads checked with python-docx and XML parsing. All six outputs reopened with docx-preview, exported as rendered pages and visually inspected. Native Word/LibreOffice fidelity was not checked: soffice.exe unavailable.
Limits: no OCR, original images or exact page design; not a general accuracy claim.
Source font: https://raw.githubusercontent.com/notofonts/noto-cjk/main/Sans/Variable/TTF/Subset/NotoSansSC-VF.ttf
Source font SHA256: d68bafcb48a2707749396aa12bbbd833cb70401f3a9a689fd2902c7e0d295964
Converter script SHA256: 2f26a2d2308554fc1b10c891f2e2b72c7d6233f3cc8f08d2b71927f7edfbd81a
