PDFMDTools PDF to Text reproducibility kit
Test date: 2026-10-13, America/Tijuana. Four self-made single-page PDFs.
Scan, columns and table reuse our PDF-to-DOCX fixtures; Chinese is a newly authored text-layer fixture.
Each input: US Letter 612 x 792 points. Columns/table: Helvetica 11pt. Chinese: Noto Sans SC fixture font, title 18pt and body 14pt. Scan: local bitmap, no text operators.
Baseline settings: Simple line order; UTF-8 BOM ON. Adjusted settings: Layout-aware text; UTF-8 BOM ON. Same source file, encoding held constant.
Before/after here means those two actual settings, not two historical deployments.
Scan: page 1 rejected in both modes, no TXT. No OCR; not repaired.
Columns: baseline LEFT01 RIGHT01 LEFT02 RIGHT02...; adjusted LEFT01-04 then RIGHT01-04. Eight identifiers once in both.
Table: baseline spaces; adjusted TAB-separated values. Seven rows including header, three fields per row; all 21 values matched.
Chinese: complete expected four lines in both modes. No failed encoding claim: the fixture passed before and after. Separate optional-BOM-off run also passed strict UTF-8 decoding.
Actual six browser downloads equal the UI text after UTF-8 BOM removal. Auxiliary mixed scan fails page 2 without partial download; rotated and encrypted inputs also rejected.
No original page design, images or rich tables; irregular layouts, merged/wrapped cells and source encoding errors need manual review.
No competing converter tested. Disclosure-gap longtail volumes pending owner input; no uniqueness or superiority claim.
Converter SHA256: 5921d8f5937e2674ae0e4bc3b73746beb76c72a75f61500e5e476681209fd851

OCR follow-up 2026-10-13: five total fixed inputs. Original scan rejections/mixed-scan rejection refer to OCR OFF. English OCR returns both exact lines. Chinese image-only OCR returns all lines but ZH003 -> ZHO03; no post-hoc correction. Tesseract.js/core 7.0.0, LSTM OEM1, PSM3, render scale <=3, longer side <=2400px, English / chi_sim+eng. Actual downloads equal preview. Disclosure demand rejected by owner: seven tested variants all zero; measured OCR/scanned phrases have demand, supplied by owner. No competitor quality absence claim.
