# JPEG compression / controlled corner test

{
  "test_date": "2026-10-19",
  "kind": "controlled synthetic test, NOT native Gemini watermarked downloads",
  "scene_seed": 20261019,
  "engine_sha256": "d956d58d3f5b195990c17549893eaf773bac4de7dba5ed4d61495f0c0f919e78",
  "alpha_map_sha256": "69fc428ed9cdf7769ca63d2d70c2b5d661acbc88adf3ff1632f882a63c0f98ed",
  "encoder": "Pillow 12.3.0; JPEG chroma subsampling=2; PNG compression=6; not browser encoding",
  "formula": "box_rgb_mae_vs_ground_truth = sum(abs(cleaned_RGB - known_original_RGB)) / (48 * 48 * 3); maximum is max of these absolute channel differences. Same-encoding-control MAE replaces known_original_RGB with decoded unmarked control encoded with identical settings. Outside count compares cleaned RGBA with decoded marked input, counting pixels with any channel changed, excluding the edit box."
}

Run reproduce.py with Python, Pillow, NumPy and Node.js available. It regenerates the synthetic scenes, encodes PNG and JPEG controls, invokes the included unchanged engine function, and writes raw results. No original Gemini downloads, browser export tests or success rates are claimed.