# Validation record

Date: **2026-10-14**.

The standalone package was tested with Python's `unittest` using synthetic PCM and mocked clients. The final run passed **15 tests**. Tests cover:

- Native PCM16 preparation without changing samples and optional reference handling.
- Refusal to reuse output directories.
- Hash checks for changed originals, references and calibration controls.
- Exact 12-second windows, half-second overlaps and gap-free coverage.
- Explicit audio-transmission and known-good-reference gates before network access.
- Clean/silent/reversed control generation and validation.
- Rejection of fabricated words in silence/reversed controls even when the model reports no intelligible speech.
- Complete mocked review with full-clip, reference and all-window passes.
- Preserving full-clip acoustic failures when all windows pass.
- Recording a network failure without an application retry.
- Rejecting malformed/missing evidence and unexpected demo revisions.
- Disabling implicit tokens and checking the Gradio client configuration.

Run from this folder:

```text
python -m unittest discover -s tests -p "test_*.py" -v
```

No real audio was sent during standalone-package development or tests. A separate read-only check of the public Hugging Face metadata on 2026-10-14 confirmed that the advertised revision and host matched the pin. This metadata check is not a live inference test.

Local CLI integration checks also passed for a synthetic PCM16 WAV and for an MP3 converted by FFmpeg into the review format. Both produced prepared bundles with no reported errors. Synthetic tones test preparation only; they are not speech calibration material.

Real Gradio audio submissions from the new package have **not** been integration-tested here. Prior successful use of the same public demo does not establish this package's live integration. Online availability, latency, limits and model behavior can change. The CLI verifies revision/host/signature again before each online operation.

The tests validate control flow and evidence integrity, not the perceptual accuracy of the model. Inspect a real calibration and first actual review before relying on the tool.