What the PDF research was designed to answer
PDF files describe positioned marks on pages. A converter must infer reading order and document structure, so a successful file read is not enough evidence of a reliable Markdown result.
The quality spike compared a structured candidate against a basic PDF.js text-item ordering baseline. It required better behavior on reading order, repeated noise, or heading structure without introducing known text loss.
Can ordinary report text be reconstructed without mixing later columns or pages into the current paragraph?
Can each Markdown section retain a reliable reference to its original PDF page?
Can the converter identify layouts where normal reading-order rules are not trustworthy?
Can image-only pages stop clearly instead of returning empty or misleading Markdown?
Tested claims and current status
The published behavior is tied to fixed fixtures, repeated conversion checks, browser tests, and a real downloadable sample. It is not a claim that every PDF layout can be reconstructed.
| Area | Status | What the evidence shows |
|---|---|---|
| Single-column reading order | Validated | Required report sections remain in source order and the tested body text stays covered. |
| Headings, paragraphs, and lists | Validated | Relative font and spacing signals rebuild basic structure with deterministic output. |
| Repeated header and footer noise | Validated | The three-page sample removed six repeated lines without deleting its required body content. |
| Page traceability | Validated | Successful output retains ordered HTML comments such as <!-- Page 1 -->. |
| Columns and table-like layouts | Warning | Affected pages remain available but are identified for comparison with the source PDF. |
| Scanned pages | Blocked | The converter returns SCANNED_OR_NO_TEXT and does not fabricate OCR output. |
How the boundaries were checked
The preflight set covered a single-column report, repeated page furniture, heading hierarchy, a two-column paper, and table or scan risk. Production tests also covered damaged input, a scan-only PDF, a risk result, and a 30-page reference file.
The same file and build must produce the same Markdown bytes, warning codes, page numbers, and ordering.
The 30-page reference completed inside the 10-second target in the tested desktop browser.
Browser tests checked the .md filename, page marker, expected text, and a real download event.
The core flow completed without sending source PDF bytes or generated Markdown to a server.
What this evidence does not prove
It does not prove high-fidelity reconstruction of every two-column paper, complex table, equation, handwritten page, embedded image meaning, or unusual PDF producer.
A warning is a request for human review, not a quality score. The original PDF remains the visual source of truth whenever layout carries meaning.