# Test Recording Outputs

The original `/workspace/testwav.wav` was not changed. These files are listening
samples, not a claim that the presets have been tuned by listening.

- `testwav.deepfilter-cleaned.wav`: full recording after DeepFilterNet3 at 85%
  wet strength, mono 48 kHz float WAV, with exactly 5,684,577 frames.
- `testwav.natural.wav`: the cleaned file through the natural mastering profile,
  48 kHz / 24-bit mono, measured at -19.11 LUFS and -1.49 dBTP.
- `testwav.natural.mp3`: 192 kb/s delivery copy, decoded measurement -19.38 LUFS
  and -1.75 dBTP.
- `testwav.fft.wav`: the original processed with the explicit non-AI FFT denoiser,
  for comparison, measured at -19.10 LUFS and -1.50 dBTP.
- JSON sidecars record settings and technical measurements for each mastering run.
- `../previews/` contains 30-second natural, narrator and radio samples starting
  at 10 seconds in the DeepFilterNet-cleaned recording, normalized to similar
  loudness for comparison.

The container uses musl rather than Arch's glibc. Real CPU DeepFilterNet inference
was tested using a temporary glibc loader; the cleaned output was then mastered
with the normal CLI and `--no-denoise`. Therefore the natural report correctly
lists the cleaned file as its source and `denoiser = none`. This two-step test did
not exercise the standard Arch installer or CUDA. It also ran denoising before
highpass filtering, whereas the normal one-command pipeline prepares/filters the
input before denoising, so a fresh Arch run is not expected to be bit-identical.

On Arch, reproduce the intended one-command workflow with the launcher:

```bash
./producer.sh process testwav.wav --device cuda --mp3 -o arch-natural.wav
./producer.sh preview testwav.wav --device cuda --start 10 --duration 30 \
  --profiles natural narrator radio --output-dir arch-previews
```

These samples were moved to `lib/archive/` so the checkout root stays clean;
the files themselves are unchanged.