# Controlled resized sparkle test

{
  "test_date": "2026-10-20",
  "measured_at_utc": "2026-10-19T19:02:29.386265+00:00",
  "scope": "Six controlled synthetic PNG scale cases. Current production gate plus unchanged inverse function in Node; not native Gemini images, actual screenshots, browser downloads, or a real-image success rate.",
  "source_files": {
    "C01-reference.png": "215c32f76459c53b92a96e106fce75bb6dfe23fd5a1ff992c49793e136ca1f97",
    "C02-reference.png": "ac76116cf822ed86c6d4e5e4fc89ea313b01c18b4a97ee2381cf6f89968299aa",
    "W01.png": "3c96e69ca29bdbf12c9393bef3981e18ee84d8bf0af2610f8f0079c9b1554374",
    "W02.png": "c812b7252a36b58807e8fee928b782c3deef0eaeadacdede0d9d60c6d128f560"
  },
  "engine_sha256": {
    "app.js": "3a2f32556201e1fbd6e700d343dffef218677611f3c9e9d8c5f6f7c9c270e085",
    "alpha-maps.js": "69fc428ed9cdf7769ca63d2d70c2b5d661acbc88adf3ff1632f882a63c0f98ed",
    "watermark-gate.js": "5029abfbed207c91b018887605c4595cb5428f0d7ef73ae95dac3ca515cf2dee"
  },
  "settings": {
    "source_width": 1024,
    "source_height": 768,
    "source_box": [
      944,
      688,
      48,
      48
    ],
    "resampling": "Pillow 12.3.0 LANCZOS",
    "scales": [
      0.75,
      1.0,
      1.25
    ],
    "input_encoding": "PNG",
    "mode": "auto",
    "offset_x": 0,
    "offset_y": 0,
    "strength": 1.0
  },
  "measurement": "Control ROI is the original 48px box scaled with floor/ceil and padded 4px on every side. MAE=sum(abs(output RGB - identically resized unmarked control RGB))/(ROI width*ROI height*3). Max is largest channel difference. Changed pixels compare pipeline output with decoded input; outside excludes the actual chosen box. Bypass writes original decoded RGBA in this isolated runner, not a test of browser original-file byte download."
}

Run `python reproduce.py --root . --node node` in the extracted package after installing Pillow and NumPy. Three engine files in site/ are byte-identical to the production snapshot. Results are synthetic, not real-image success rates.