StemGround: Source-Interventional Auditing and Selective Response Alignment for Instrument Grounding in Music Foundation Model Readouts

Abstract An accurate instrument tag can persist after the instrument is removed, exposing a distinction between recognition and source grounding. We introduce StemGround to measure this distinction and train source-aligned readouts using ground-truth multitrack recordings. Same-recording target removal, ordered gain changes, and target-by-context interventions specify which outputs should change and which label truths should be preserved. Loudness matching, whole-mixture gain, and unrelated-source controls test acoustic alternatives. Selective response alignment (SRA) implements these require- ments through a 263,171-parameter residual head, leaving the music encoder fixed. We audit MERT, MuQ, MusicFM, and a MuQ-MuLan zero-shot reference, then compare SRA with an ordinary head matched for capacity, input data, and selected training updates. Both heads receive labeled interven- tions. Improvements on artist-disjoint MoisesDB evaluation replicate on the frozen MUSDB18 official test: 50 tracks from 46 artists, with 13,361 interventions across 49 eligible tracks. Relative to ordinary training, SRA increases standardized target-removal efect by 0.664 and dose-order consistency by 0.043. It reduces target-absent false positives by 0.161, net context-induced false positives by 0.300, and standardized target-absent context efect by 0.756. All five endpoint-wise 95% artist-cluster con- fidence intervals favor SRA; cross-model natural-task and of-target noninferiority criteria are met. MusicFM exhibits a target-removal trade-of, and probability reliability worsens. StemGround estab- lishes a measurable and trainable connection between physical source changes and instrument-label responses, distinct from natural recognition accuracy and probability calibration.