music/records/benchmark/PROBE_DEMUCS_VOCALS.licence.json
{
"canon_id": "PROBE_DEMUCS_VOCALS",
"registry": "",
"lane": "music",
"generating_model": "demucs-htdemucs",
"model_revision": "htdemucs",
"model_licence": "MIT",
"licence_class": "previz_only",
"previz_only": true,
"prompt_hash": "27abcde80e1c45b6303d4a98d3114ad14de5419c8bb91d6014e7b3925c8857f4",
"seed": null,
"generation_tier": "analysis",
"decay_stage": "n_a",
"presentation_tier_ref": "",
"render_tier": "ANALYSIS",
"style_lora_hash": "",
"out_path": "build/audio/benchmark/stems/demucs/htdemucs/PROBE_MIX/vocals.wav",
"sha256": "bc97d23dbe5f3a60b2e2d4d24cd357669bd2f4c87cc9c06ccc6f74d80e300d1d",
"bytes": 4233644,
"notes": "DERIVED ANALYSIS ARTIFACT: the 'vocals' stem Demucs htdemucs separated out of PROBE_MIX.wav (sha256 9dd5d9fdc6b06d63...). Model licence MIT, re-read from https://github.com/adefossez/demucs/blob/main/LICENSE. It exists to cross-check ACE-Step's `extract` against a real separator (benchmark M3) and never ships, which is what licence_class=previz_only records. prompt_hash here is a DERIVATION hash over the derivation block, NOT a generation_prompt_hash -- a separator has no prompt, and the two must never be compared. No Sonniss byte is an input to this artifact: the source is our own generated probe mixdown."
}