How the pulled test files converted

Success by conversion path

Faithful enough diff checks

These checks compare generated outputs against Haskell Pandoc when it can read the source, a format-native external reference where one is available, or PHP HTML only as a disclosed fallback. Text scores compare normalized visible words. Visual-structure scores compare the rendered document shape: headings, paragraphs, lists, tables, images, figures, captions, code, quotes, and math. Generic XML uses an independent libxml2/libxslt transform over common structural element names. PDFKit supplies text and geometry, not an HTML semantic tree, so PDF visual tag scores are explicitly excluded rather than reported as a false mismatch.

Text

Visual structure

Bibliography semantics

Standalone BibTeX, BibLaTeX, CSL JSON, EndNote XML, and RIS inputs are rendered by Haskell Pandoc through Citeproc. Their WordPress output is checked by exact reference count and Citeproc token coverage, not by wrapper tags, because the PHP port deliberately emits editable definition-list blocks.

Import quality gates

These checks evaluate the WordPress block output as an import artifact: visible text completeness, paragraph merge or split drift, Citeproc reference coverage where applicable, media extraction diagnostics, local anchor validity, and Custom HTML block share. Generic XML is checked against independent libxml2/libxslt source structure. PDFs additionally use independent native source-token coverage and line geometry because untagged PDFs do not encode HTML heading, list, table, or image semantics.

Actionable thresholds

The common-format gate is blocking and covers normal WordPress import formats. Exotic formats are tracked separately so they stay visible without diluting the common-format release signal.

Curated stress showcase

These are representative real-world files from the pulled corpus: leaflets, brochures, a scanned book, image-heavy packages, table-heavy office documents, and rich markup fixtures.

Extracted media preview

The PHP path now runs an --extract-media-style pass in images we thought were important mode. Referenced package images are written beside the converted output and their <img> URLs are rewritten to hosted files; directly embeddable PDF streams are copied out, while JBIG2 streams are decoded through the bundled browser-compatible PDF.js/PDFium raster path into validated 1-bit PNG media.

222 image media entries extracted across 21 source files.

Success by input format

WordPress block coverage

The WordPress output generated blocks for 108 samples.