# Article 6 benchmark evidence

This directory contains the reproducible evidence for the unpublished CompressTo article “500KB vs 200KB vs 100KB vs 50KB: How Much Image Quality Do You Actually Lose?”

## Controlled environment

- Test date: 8 August 2026
- Production URL: `https://compressto.tools/`
- Browser: Chromium 151.0.7922.76 in the Codex in-app browser
- Operating system: Windows
- Production worker identifier: `compressor.worker-Bv-t9fAi.js`
- Repository commit at test time: `a2e9d3cc60525f1a33a78d6df2a600605956ec77`
- Output: JPEG
- Preserve dimensions: off
- Size calculation: bytes ÷ 1,024

## Source-size reconciliation

`real-photo-original.png` is exactly **1,734,688 bytes**, or **1694.03125KB** when bytes are divided by 1,024. Its SHA-256 checksum is `27257fb3fa87a148fcf05d51ef9f1e562dedb883dff2ea42f954b0cb2a668821`. The earlier manifest display of `1694.031KB` was the same byte count rounded to three decimal places, not a different source file.

The four photograph reductions recalculated directly from 1,734,688 bytes are 85.47756138279621465070375767861425224594% (500KB ceiling), 88.38298299175413676695751627958457082772% (200KB), 94.63131122138390304193030677562766330314% (100KB), and 97.12415143241897101957239572764670073235% (50KB). Exact bytes are canonical; shortened percentages in publication tables are explicitly marked with an ellipsis.

All 12 outputs were generated from the untouched original retained by the production CompressTo interface. The interface’s “Try a different target size” control explicitly recompresses that original, not the prior output.

## Sources

- `real-photo-original.png` is the genuine project-owned desktop-workstation photograph supplied for this benchmark. It is the untouched 1086×1448 PNG used as the source for all four photo runs.
- `source/fictional-small-text-document.png` and `source/fictional-product-graphic.png` are deterministic project-owned images created by `scripts/generate-article6-sources.py`.
- Every original exceeds 512,000 bytes because of its visual content and resolution. No padding bytes or hidden payloads were added.

Recreate the deterministic sources with:

```powershell
& 'C:\Users\pc\.cache\codex-runtimes\codex-primary-runtime\dependencies\python\python.exe' '.\scripts\generate-article6-sources.py'
```

The source generator intentionally creates only the document and product graphic. The genuine photograph is retained unchanged at the article-package root.

## Publication privacy review

The project-owned workstation photograph and every publication PNG/WebP derived from it were inspected at native resolution. No password, API key, personal information, private URL or confidential client material is legible. The monitor contains only tiny generic code-like text; incidental product branding is visible and is acceptable for this benchmark. The verified images were not altered for branding removal.

## Measurements and graphics

- `results-manifest.json` records the environment, exact bytes, decoded dimensions, reductions and SHA-256 checksums.
- `results-manifest.csv` provides the 12 output rows in a portable table.
- `graphics/*-comparison.*` fits the original and four outputs into identical display boxes in the fixed order Original, 500KB, 200KB, 100KB and 50KB.
- `graphics/*-100-percent-crop.*` stacks the same source region at native pixels without resampling. When CompressTo resized an output, coordinates are mapped proportionally and the smaller native crop remains smaller.
- `graphics/crops/*-crop-coordinates.json` records every crop box and mapping rule.

Rebuild only the genuine-photo comparison branch, combined hero and manifests with:

```powershell
& 'C:\Users\pc\.cache\codex-runtimes\codex-primary-runtime\dependencies\python\python.exe' '.\scripts\build-article6-evidence.py' --subject real-photo
```

## Run note

The product graphic’s 50KB compression completed and the production page displayed a 49.2KB result. Browser automation timed out while dispatching the ordinary download button, so the current compressed preview’s JPEG blob was saved directly through the browser’s media-download capability. Its 50,423-byte file is the same production result shown by the page; it was not recompressed or substituted.

The article remains unpublished. The Sanity document, guarded publisher, guarded inbound-link updater, schema metadata and sitemap fixture are prepared in the repository; their write operations must only be performed after editorial approval.