TRACE companion — asset credits, licenses, and code provenance

Paper: Voice Agents Under Acoustic Stress: From Signal Degradation to Interaction
and Action. Amir Ivry, Kai-Wei Chang, Lin Zhang, Sharon Gannot, Carlos Busso.
Submitted to ICASSP 2027; manuscript dated 2026. The included PDF is unchanged.
No separate paper reuse license was supplied; no new license is asserted here.

Speech: Loren Lugosch, Timers and Such v1.0, dev-real/anonymous0_24.wav.
https://zenodo.org/records/4623772
Source SHA-256: c303786d4f544d6afa845bceeecadede2f60b717308d5a60afd51650f272d3df
CC0 1.0 Universal. Full text: source/TRACE_COMPANION_TRANSFER/assets/speech/CC0_1.0.txt
(in the webpage preview: licenses/speech_CC0_1.0.txt).

Machinery: ldezem / Ore.e Refineries, Sound of Work: Blacksmith edition 1.
Pneumatic chipping hammer — Holman nc4 — Start 1 short.
Recorded at Jesse Sipola's smithy near Turku, Finland.
https://oree.storijapan.net/praxis/sound-of-work/
https://freesound.org/people/ldezem/sounds/386239/
Source SHA-256: 6714b1bc2d4d3ca4e340387c9295ee2ef75296106451d5a1b14999e3236a9abc
CC0 1.0 on Freesound; creator metadata marks PD. The selected released recording
has creator-applied mastering and beat-length editing; it is not raw ambience.
Full text: source/TRACE_COMPANION_TRANSFER/assets/noise/CC0_1.0.txt
(in the webpage preview: licenses/noise_CC0_1.0.txt).

Companion audio derivatives use one shared gain and PCM16 quantization. The
machinery derivative mixes the full released event. The silenced derivative
zeros a historically aligned interval. The untouched source assets remain included.

Historical model: Qwen team, Qwen/Qwen3-Omni-30B-A3B-Instruct.
https://huggingface.co/Qwen/Qwen3-Omni-30B-A3B-Instruct
Revision: 26291f793822fb6be9555850f06dfe95f2d7e695
One archived response from 9 October 2026 remains separate from the new gallery
observations. New saved first responses, where completed, are described in
RESULTS_GALLERY.md. No weights, installed environments, credentials, or live timer
services are bundled. Optional inference source/configuration is included.

Code provenance: the transferred audio_recipe.py was extracted/re-expressed from
the historical v3 prepare.py additive operation. The companion native_parser.py
derives from the transferred parser, with the documented malformed-tail fix in
companion/PARSER_CHANGE.md. Source hashes and adaptation provenance are provided
in source/TRACE_COMPANION_TRANSFER/configs/code_provenance.json. The conversation
scorer, saved-response validator, site, build/export tools, and new tests were
added for this companion. CPU tools remain separate from optional GPU code.
Audio CC0 licenses do not license the code.
No explicit code license was supplied for the transfer or companion; none is
invented for this download. Redistribution/reuse licensing remains with the
rights holders. No blanket license is asserted for the paper or entire archive.

The public code archive is an allowlisted subset of the transfer. It deliberately
omits private handoff instructions, old environment/runtime receipts, original
historical full traces containing host paths, private scheduler receipts, caches,
and weights. The optional unchanged inference entry point and its required
configurations are included with GPU_REPRODUCTION.md; CPU dependencies remain separate.
evidence/historical.json is a labeled public extract retaining exact response,
prompt, tokens, settings, and recorded simulation, plus the complete source-trace
hash. The immutable original trace remains in the local research project.
