Original compression and saved prediction study

Native codec files use FFmpeg 9.0.1 with libx264 and FFV1
The Python builder uses NumPy 2.3.5 and Pillow 12.0.0
In a fresh output folder run the published builder and then verifier
uv run --no-project --with numpy==2.3.5 --with pillow==12.0.0 python build-compression-study.py OUTPUT
uv run --no-project --with numpy==2.3.5 --with pillow==12.0.0 python verify-compression-study.py OUTPUT
The native builder also saves an intermediate raw YUV file for reconstruction
Published measurements and crops use every decoded luma sample with a fixed 640 by 360 raster and 72 frames at 24 fps
The original source decoder must return byte identical raw YUV to the authored stream
Native container bytes can vary with encoder version CPU and Matroska identifiers so compare decoded files and numeric records before container hashes
The PNG crops map stored limited range luma bytes directly to grayscale bytes and are not a full RGB playback simulation

Saved record review uses Bun 1.3.11 and sharp 0.34.5
Copy the helper and artwork builder plus native media records source audit validation scripts and this text into the output folder
bun add sharp@0.34.5
bun add --dev @biomejs/biome@2.2.2
The artwork builder formats generated JSON with that pinned Biome CLI before hashing
bun build-artifact-review-assets.ts OUTPUT
The default JSON record is authored numeric data rather than a model response
The source validation script fetches one hash pinned stdlib source module or accepts its local path
python3 verify-verus-source.py PATH_TO_COMMON_PY
That test covers only selected JSON and completion functions with no weights or GPU inference
It does not test model quality detect video origin or establish semantic correctness of a saved prediction
The normalized source coordinates map to user supplied frame dimensions and the SVG overlays all supplied boxes on a blank frame
All source model and dataset licenses remain distinct from this original fixture