Rachel

Incrementally generated unfiltered candidates. This is not a final selected dataset. Each background has 15 separately generated prompt–seed outputs using gain-compensated reconstruction residuals. runconfig.json pins models, source pools, parameters and implementation hashes. For multiple workers read workers/worker-NN/progress.json; each worker reports only its assigned IDs. Global completion requires all worker reports to be complete. Assignments are backgroundid modulo numworkers. Per-worker runconfig.json pins the parameters. Workers never share a background ID. Each data/group-NNN/noise-NNNNNN.tar contains one normalized background.wav, 15 rank-NN.mask.wav / rank-NN.mixture.wav pairs…