# Release validation

The public adapter SHA-256 is
`c26031b9e1932ae0e76b29ec894b49bb6f719a9ff997cf733eae8bbf443f9fb2`.
The loader refuses a different adapter checksum, parameter inventory or shape.

## CPU checks — seven passed

- Exact adapter SHA and 292 FP32 tensor inventory.
- Rank-128 A/B pairs totaling 96,337,920 scalars.
- Full 146 attention-projection base skeleton on the meta device: every inserted
  target name/shape matches the real saved adapter, without allocating model weights.
- Zero-initialized residual and FP32 adapter storage.
- Finite component-normalized advantages with invalid candidates excluded.
- Warmup/cosine schedule endpoints.
- Finite VESPO policy-loss gradients.

## Allocated-GPU portability canary — four passed

The **new portable inference loader**, pinned local base copy, exact adapter and
external codec were exercised on four GH200 GPUs under Slurm. Two short English
and two short German generations cover both reference/no-reference paths.
Each produced finite 48-kHz mono audio with 12 valid codec indices per frame.
Outputs were kept in four SQLite containers, not loose audio files.

The tested runtime was PyTorch 2.9.1+cu128, CUDA 12.8 and Transformers 5.14.1.
The base loader uses BF16 SDPA; custom adapter multiplication remains FP32 at
scale 1.0. A generic upstream missing-key warning concerns explicitly tied text
and codebook output heads (the base implementation ties them to embeddings),
not missing adapter tensors; all 292 adapter tensor names/shapes are checked.

This is a functional loading/generation test, **not** a subjective quality
evaluation or proof of voice identity. The separately completed 70-case matched
listening evaluation provides the aggregate quality numbers in the model card.
No fresh GPU training smoke is claimed for the public portable trainer's
additional warm-start argument; original scientific core files are unchanged,
while portability differences are documented in TRAINING.md.

See [machine-readable validation](stats/validation.json).