# Validation

- Dataset pinned to `32074728eed590052b31f23d1df2a6ce92ada94f`.
- All 150 local scene data.h5 SHA-256 values matched the dataset's published LFS SHA-256 values during preparation.
- Both methods cover exactly the same 150 dataset folders; 300 nonempty weights total.
- All 300 files loaded with `torch.load(..., weights_only=True, map_location="cpu")`.
- All files have the same 600 tensor keys/shapes, 20,152,320 parameters, float32 storage, rank 8, and finite values.
- Source-to-package SHA-256 matched for every weight; all DanceGRPO download checksums also matched the remote source.
- All 150 DanceGRPO soups exactly equal the element-wise mean of their eight source epoch checkpoints (maximum absolute difference zero).
- All 150 DiffusionNFT source training logs record saving epochs 0–7 and update geometry of six windows, 24 rollouts, and 48 draws. Three seed videos are present for each source scene.

DiffusionNFT soup arithmetic was not independently recomputed: the selected local source contains the final soup and logs, but not the eight source epoch files. The original soup files are preserved byte-for-byte. This package verification does not rerun GPU inference or establish numerical reproduction of every manuscript result.

Source files and validation evidence were checked before publication.