Check source and record run facts
Hash bytes; enforce file and page limits; inspect encryption, boxes, rotation, and damage; record tools and options.
Level 2 · System flow
Last modified: (dates/navigation repair). Content current as of: (last retained content commit; content verification date unavailable). Original content last edited: . Current claims have not been freshly reverified.
From PDF input to a document record
The conversion route preserves observations before interpretation. It then builds semantic candidates, writes PDFMD, parses the result again, validates it, and only then renders Markdown.
Main route
Canonical PDFMD sits between recognition and rendering
PDFsource bytes
→
Evidence bundleobservations
→
Semantic draftassertions + alternatives
→
PDFMDcanonical record
→
Markdownderived view
Markdown is never reparsed to recover the canonical record. The renderer consumes PDFMD and its package assets directly.
Dependency order
Eight processing stages
Code
Collect observations
Capture native glyphs, OCR tokens, layout boxes, paths, annotations, and images in normalized page coordinates.
Propose lines, columns, and order
Group baselines and score reading-order candidates using geometry and local continuity.
Identify page furniture and number systems
Use recurrence and sequence evidence so headers and footers do not become body text.
Resolve inline appearance
Attach bold, italic, decoration, scripts, color, and highlights to text runs.
Propose page-local structure
Headings, paragraphs, lists, quotations, notes, tables, and graphics.
Resolve cross-page structure
Join paragraph, list-item, footnote, and table fragments across page events.
Emit, reparse, validate, render
Write PDFMD, parse it, validate references and assets, then generate the human view.
Input routes
Native text and scanned text
Native PDF
Decoded glyphs, text matrices, baselines, fonts, paint operations, paths, and annotations are direct observations. OCR may remain as disagreement evidence but cannot replace good native text without a recorded reason.
Scanned or raster region
The page or selected image region is rendered at a recorded resolution. Tesseract supplies tokens and boxes. Missing OCR remains an explicit issue, and the raster asset is retained rather than silently dropped.
Current implementation
Implemented route; incomplete recognition scope
The capture → OCR → semantic draft → PDFMD → parse → validate → Markdown route is implemented. Recognition methods exist for the main element classes, but their accepted reliability differs. Complete paragraph assembly, exact scanned-font identification, a general resolver, model routing, and interactive review remain incomplete or planned.
Do not equate “valid” with “correct.” Format validation proves that the PDFMD structure obeys its contract. Source fidelity requires independent expected results or full page-by-page review.