VALIDATION.md

v0.1.0 preview validation

The public release files were checked before publication and again through an anonymous download from the public Hugging Face model repo.

• The published acoustic file has SHA-256 ef80763236b2467a886d52fba51769de4dcfbdce909dd320803b6d2d2d41db96, 47 tensors, and 54,294,272 parameters. The adapter SHA-256 is e2b53154b40cd67faf3c9a57226f09c187b060569a7894ee2b3630a4e88937c3. The base is pinned to google/gemma-4-E2B-it@3e22461f65e89153144f8adb70e3b8c2cc9845a7 and checked by the loader.
• The release loader ran on an NVIDIA RTX PRO 6000 Blackwell Server Edition MIG 1g.24gb with PyTorch 2.13.0+cu130, Transformers 5.17.0, and PEFT 0.21.0. It returned valid choice distributions for three named questions on a short CREMA-D speech sample and a synthetic 58-second two-chunk input. No sample audio is bundled.
• The public-Hub load gave exactly the same three answer distributions as the same release files loaded from a local staging directory. PyTorch reported 9,773.9 MiB allocated. The simple sequential loader took 1.778 seconds to score three short-input questions on the public-Hub repeat; the synthetic 58-second local-file smoke took 3.669 seconds for three questions. These are functional checks under different cache conditions, not warm-service latency measurements.
• The GPU pod was deleted after testing. The optimized shared-audio HTTP timings in EVALUATION.md come from a separate staged serving path and should not be attributed to this reference loader.

The GitHub repository contains only code, docs, examples, configuration, and hashes. Hugging Face contains the adapter and acoustic delta plus the same documentation. Neither contains training/evaluation audio, the base model weights, optimizer state, credentials, or private row-level predictions.