What text does a resume PDF actually give a parser?

We generated six fictional resumes, ran the same native text-layer parser against every file, and published the fixtures, source, and exact output. The test isolates text extraction; it does not score resume quality.

What this run shows

• All anchors were returned in order for the single-column, two-column, Korean, mixed-language, and table/link fixtures.
• The table/link fixture returned its visible table cells, URL, and link label. This parser does not expose the destination behind a labeled PDF link, so destination preservation was not evaluated.
• The image-only negative control returned no native text. We did not run OCR, so this result says nothing about OCR accuracy.

Method and limits

The six one-page fixtures contain only fictional names and organizations. They were generated with ReportLab 4.4.9 and Pillow 12.3.0, then measured with the default text extraction in pdf-parse 1.1.1. The script stores the complete extracted text, anchor counts and order, page count, byte size, and SHA-256 hash in the result JSON.

The baseline JSON is an archival record of the September 11 run. The original two-column and image-only binaries are not bundled, so those two historical hashes and byte counts cannot be reproduced from the current generator.

This is not an ATS acceptance test, an OCR experiment, a resume-field accuracy test, or a test of the refresh.cv production importer. A successful extraction in these controlled files does not predict screening or hiring. Review imported names, dates, links, sections, and bullets against the source PDF.