# Synthetic resume PDF native-text experiment

This experiment creates six fictional, one-page resume PDFs and measures the
text returned by `pdf-parse` 1.1.1. The fixtures cover single-column and
two-column layouts, Korean text, mixed Korean and English text, a table with
links, and an image-only scan used as a native-text negative control. They
contain no personal resume data.

## Reproduce

From the website root:

```sh
npm install
python3 -m venv /tmp/refresh-pdf-study-venv
. /tmp/refresh-pdf-study-venv/bin/activate
python -m pip install reportlab==4.4.9 pillow==12.3.0
python scripts/pdf-extraction-study/generate_fixtures.py
node scripts/pdf-extraction-study/measure-native-text.mjs
```

The Korean fixtures use the included subset of Noto Sans KR under the SIL Open
Font License. Keep the TTF beside the generator. Set `PDF_STUDY_OUTPUT_DIR` for
a different output folder.

The generator writes the PDFs and a public copy of itself. The measurement
script writes exact extracted text, expected and observed anchors, file sizes,
page counts, SHA-256 hashes, and a public copy of itself to
`public/research/pdf-extraction/`.

`native-text-results-v1.json` is an archival record of the original
three-fixture run from 2026-09-11. The original two-column and image-only PDF
binaries are not bundled; their historical hashes and byte counts cannot be
reproduced from the current generator. `native-text-results.json` is the
reproducible current six-fixture run.

## Scope

This is a native text-layer test. It does not run OCR, score ATS compatibility,
infer resume fields, test hiring outcomes, or exercise the refresh.cv production
importer. The table/link fixture measures visible text only. It contains two PDF
link annotations, but `pdf-parse` does not report their destination metadata.