v0.1.0 - Initial publication

Reproducible Pareto curve across three speech-codec classes.

Headline results

Channel	Where it ships	Reliable rate
Opus 48k AUDIO	Discord, music-mode WebRTC	4000 bps modem-tone, 933 bps real-speech-cover
Opus 24k VOIP	Zoom, Teams, WebRTC	3196 bps modem-tone, 270 bps real-speech-cover
AMR-NB 12.2k	2G/3G cellular voice	76 bps real-speech-cover

All post-FEC, multi-seed validated through real ffmpeg codecs (not surrogates). Zero residual errors observed on 76,000+ bits per headline checkpoint.

What is included

• Six trained checkpoints under core/neural_codec/ totalling ~109 MB
• Hand-coded TRIZ pitch modem (175 bps reliable, no neural network needed)
• End-to-end voice-within-voice demos for each codec
• Multi-probe synthesis comparing EnCodec, Mimi, DAC token survival through Opus
• Stacked stego_opus + prosody channel (538 bps reliable, demoed with Codec2 700C)
• Listenable WAV samples for every variant

What this is not

• Not a commercial product. Productization analysis in docs/FINDINGS_AND_PLAN.md.
• Not novel methodology. Methods are published prior art.
• Not above the published state of the art for AMR-NB stego (codec-internal QIM methods report 1-3 kbps).

License

Apache License 2.0. The patent-grant clause matters here because audio watermarking is a patent-active field.