Validation

Evaluated checkpoint

The restored recipe is the three-encoder, 8,000-reference-video run. The original checkpoint header records step 80, 1,095 generator parameter tensors, and all three branches: DINOv3, V-JEPA 2, and VideoMAEv2. Its bank identity matches the released training config. The historical human-evaluation generation records and the VLM source-video run point to this same checkpoint.

The original step-80 log reports loss 3.854, generator gradient norm 10.98, three encoder VJPs, batch geometry 256 / 128 / 128, zero reported feature replay error for all three encoders, and a completed optimizer update. These are historical measurements, not a new run of this package.

Checks of the corrected release

51 CPU tests passed; 1 CUDA-only test skipped. Coverage includes three-encoder configuration, bank loading, MMD gradients, replay/offload, the 200-update cosine curve, DINOv3 metadata relocation, terminal cleanup and context-cache inference, and five-sample noise ordering.
All 184 fields from the original training configuration were accounted for. Asset paths and descriptive identities are public equivalents; the public stopping point is 80 rather than 200, while the cosine horizon remains 200. The three active encoders are specified explicitly.
The public sampler was compared with the actual historical evaluation sampler using a small recording generator and the real scheduler. Outputs and every generator input/timestep matched exactly for two consecutive prompts, both with ordinary random draws and with supplied noise tensors.
The five-sample adapter draws the same 21 main transition tensors and 14 auxiliary tensors per prompt as the original 81-frame inference adapter, before single-video execution. CPU tests verify noise slices and final RNG state.
DINOv3 metadata relocation reproduces the complete actual frozen-bank metadata and digest without loading models from the original paths.
The single-encoder MMD kernel, training rollout, and encoder VJP implementation match the original three-encoder snapshot's executable AST after removing documentation and restoring readability aliases.
Python parsing, shell syntax, and diff whitespace checks passed.

Full 14B GPU training, GPU video generation, and VBench were not rerun for this correction. CPU sampler comparisons use a small test generator, not the 14B network, and do not establish identical generated videos or a new benchmark score.

Run the CPU suite from 14B/:

python scripts/test_cpu.py