StreamGen accompaniment
Question: does training stream-music-gen (Wu et al., 2025) on SPDMX multitrack improve over the Slakh baseline under matched steps?
Pilot config: dec_online, future_visibility = 0.
Arms: Slakh | SPDMX multitrack | optional union. Eval: Slakh-test COCOLA / Beat F1 / FAD;
secondary SPDMX held-out. Domain gap is a known risk (as in Demucs).
Metrics and audio examples will land here when the pilot finishes.
Scaffold: experiments/streamgen/.