# PDF compression study - 21 October 2026 Local synthetic measurements using the iHatePDF compression handler. All fixtures are generated by code; none contains a real passport, customer file or other personal data. The coloured scene is synthetic geometry, not a real photograph. This is not a competitor comparison or a production SLA. Reproduction from the project root (project Python dependencies and Ghostscript required): PYTHONPATH=backend backend/.venv/bin/python scripts/seo_compression_study.py --output frontend/public/research/pdf-compression-2026-10 The script is included as reproduce.py. It requires the matching project backend and is not a standalone web-service client. A fixed seed and fixed PDF metadata generate identical input SHA-256 within the recorded environment. Repeated output hashes can differ because PDF writers generate identifiers. The manifest records the source commit, script hash and software versions. Results: results.csv (summary), results.json (parameters, hashes and handler diagnostics). Every PDF was reopened and every page rendered at 72 dpi; page count, dimensions and extracted text were compared. These checks do not prove semantic correctness or legibility of every detail. The *.png files compare page one before and after. Inspect detailed regions at your intended display or print resolution. The mean RGB difference is a pixel diagnostic, not a readability score. Units: exact bytes in tables; target MB uses the handler's 1,048,576 bytes. A destination using decimal MB may need a smaller target. Duration is measured locally and must not be advertised as service speed. No guarantee of a fixed compression ratio: the intentionally unoptimised text fixture and noisy artificial scans have unusually high headroom. Already compressed real documents may barely shrink. The target modes can stop after light optimisation if the input already fits. They are not always smaller than balance mode, because the target is a size ceiling rather than maximal compression. Fixture license: CC0-1.0. Synthetic inputs may be reused and attributed voluntarily. Compression results and tabular measurements may also be reused. Third-party engine licenses remain their own. Input files and output folders: - text.pdf; text-balance/compressed.pdf; text-target-2mb/compressed.pdf; text-target-5mb/compressed.pdf - color-scan.pdf; color-scan-balance/compressed.pdf; color-scan-target-2mb/compressed.pdf; color-scan-target-5mb/compressed.pdf - grayscale-scan.pdf; grayscale-scan-balance/compressed.pdf; grayscale-scan-target-2mb/compressed.pdf; grayscale-scan-target-5mb/compressed.pdf - synthetic-photo.pdf; synthetic-photo-balance/compressed.pdf; synthetic-photo-target-2mb/compressed.pdf; synthetic-photo-target-5mb/compressed.pdf - optimized.pdf; optimized-balance/compressed.pdf; optimized-target-2mb/compressed.pdf; optimized-target-5mb/compressed.pdf