# Screenshot export study — protocol fixed before capture

Three owned synthetic browser fixtures: a light UI table, a dark code editor, and a color/texture-heavy visual report. None contains real customer data or represents a photograph corpus.

Capture each twice in fresh Chrome contexts: viewport1600×900 CSS px, device scale2, fullpagefalse, animations/caret disabled, network restricted to localhost. Expected3200×1800 pixels. Capture hashes must match across repeats before export comparison.

Export independently from each original at widths3200,1600,800, preserving aspect ratio. Pillow LANCZOS resampling; RGB opaque images. Five exports: optimized PNG; JPEGq60 andq85 with4:4:4 chroma; lossyWebPq85 method6; losslessWebP method6. The same numeric quality across codecs is not equal visual quality. No encoding-time benchmark.

90 outputs=3fixtures×2capture repetitions×3widths×5exports. Retain source hashes, output bytes/dimensions/hashes and pixel equality against the resized uncompressed reference at each width. Pixel equality is not a reading-accuracy score. Repetition checks determinism, not independent generalization.

Inspect known small-text regions at1:1 and enlarged nearest-neighbor crops. Report visual observations as bounded examples, not universal readability thresholds or OCR accuracy. Do not claim results for Preview's quality slider, chat recompression, photography, transparency, color-profile retention, macOS keyboard screenshots, third-party capture apps, or AI recognition. Apple workflow instructions are separately documented, not benchmarked here.