# ResumeGet Resume Parsing Lab - July 2026

## Scope

This release measures text extraction by the ResumeGet production PDF and DOCX upload pipeline. It does not test, certify, rank, or represent any named applicant tracking system.

## Changes from 2026-07-v1

- Corrected PDF name and contact line spacing across nine affected fixture layouts.
- Added automated text-collision validation for every generated resume PDF.

## Corpus

- 10 synthetic resume layouts
- 20 PDF and DOCX artifacts
- 11 extraction-risk patterns
- no real person, employer, institution, account, upload, or application data

Every person, employer, institution, URL, phone number, and achievement in this release is synthetic and reserved for testing.

## Execution

Each generated artifact is passed to `extractTextFromUpload`, the same extraction function used by ResumeGet's production import route. The run records the extracted text, its SHA-256 checksum, the file checksum, parser errors, expected marker presence, and marker order.

## Result labels

- **Preserved:** every expected marker is present in the expected order.
- **Reordered:** every expected marker is present, but at least one appears outside the expected reading order.
- **Content loss:** extraction completes, but one or more expected markers are absent.
- **Unreadable:** the production extractor rejects the artifact or returns fewer than 120 characters.

The classifier checks a bounded set of disclosed markers. A preserved result does not prove that every visual relationship survived, and none of these labels predicts how a third-party hiring system behaves.

## Reproduction

```bash
pnpm install --frozen-lockfile
pnpm lab:generate
pnpm lab:validate
```

The canonical case declaration is `research/parsing-lab/cases.json`. Generated artifacts, raw extracted text, the result table, manifest, and checksum list live under `public/research/resume-parsing-lab/2026-07-v2/`.

## Limitations

This corpus is intentionally small, synthetic, English-authoritative, and focused on deterministic text extraction. It does not score resume quality, semantic understanding, OCR, hiring outcomes, visual fidelity, accessibility of authoring software, or compatibility with external products. Results are observations for release `2026-07-v2`, not universal claims.