Measured Influence: Pre-registered, Power-gated Validation and Calibration of Per-output Training-data Attribution for Consented Music Adapters

Abstract

When a generative music model is adapted to one artist’s consented catalogue, the rights in each output can be computed from how much each input track influenced it and who holds the rights in each input track. The second is a record; the first must be measured, per output, with a stated error. We report a pre-registered, power-gated validation of such a measurement on low-rank adapters (LoKr, rank 32) to a 2.4 B-parameter flow-matching audio model (ACE-Step 1.5), against influence by subset retraining as ground truth: 220 subset adapters on a twelve-input-track catalogue and 165 on a second. Every confirmatory test had its bars committed before scoring, and every per-output test had to demonstrate the ground truth’s own resolving power before a verdict was read from it. On ordering, the D-TRAK kernel with velocity-norm features reaches a linear datamodeling score of 0.70 (0.67 to 0.74) over twenty fixed half-size subsets, above cosine (0.46) and a size-only prediction (0.38) on 92% of outputs. On per-output magnitude, the quantity a payment requires, we define the share of a unit payment that lands on the same tracks under the method’s split and the ground truth’s (one minus their total variation distance), set a bar of 0.70, and find 0.802 (0.797 to 0.807) with calibration slope 1.00 against 0.695 for a split by training-data share. On the second catalogue the first reading drew no verdict because one of 165 adapters had diverged in training and the instrument failed its power gate; under a run-health rule committed before re-scoring, it passed at 0.749 on tracks and 0.916 on the right writers under invented co-writer splits. A calibrated split, the kernel’s share times one fixed factor per track fitted once on the retraining record, then passed its own pre-registered tests on fresh outputs at new seeds: 0.879 on the first catalogue, 0.874 on the second, and 0.900 on the production path. Four further tests are reported with their outcomes: dose response, whose expected direction retraining showed to be wrong; delivery invariance, where shares move under 0.05 but the first-named track changes on about one output in ten; known answers, failed three times; and abstention, failed on an embedding rule and passed on a loss-margin rule. An independent unlearning-based method failed by construction of its registered rule. A float16 fault that had reversed the sign of a baseline is disclosed. Results are for two catalogues by one composer; nothing here is evidence of generality.

Key results

Figures as given in the abstract, with the confidence intervals the abstract states.

The full record of one generated piece, with its influence shares, splits and signed export, is published here.

What this paper does not show

The paper’s Limitations section, in full.

• Two catalogues, one composer. Neither catalogue has real co-writers. The split only moves money between different rights holders, and that has been tested once, on invented splits. A catalogue with real co-written splits by another composer is the next proof, and with it a larger catalogue that can be read at 15, 25 and 40 tracks.
• Ground truth is retraining loss. “Correct” throughout means the split a retraining ground truth assigns to an output’s velocity-matching loss, which is the standard datamodel criterion and not a listener’s judgement. The one listening record (T10) was labelled.
• The run-health rule. It was written after the diverged run had been found by looking. Its first independent test is the next catalogue.
• Known answers fail. Under the meaning of credit, retraining names a training section’s own track first and the kernel often does not; no readout tried has cleared the bar, and the calibration does not address it.
• The first-named track is unstable on about one output in ten under the scoring’s own noise, though shares move under 0.05.
• Repetition. A gradient kernel taken after training under-credits material the adapter has learned well; the direction of its response to repetition matches retraining, its size does not.
• Partial informativeness and rank. 0.55 beyond the pooled ordering on the research index; rank agreement under 0.7 on both catalogues except on one fresh-output run.
• Served path on nine tracks; second catalogue’s served path unrun. The calibrated split is served for the second catalogue on the strength of the first’s served reading.
• Single-model index. No ensembling; λ chosen on three held-out tracks and consistently beaten on outputs by a smaller value not pre-registered; ten fixed timesteps unmeasured; per-timestep renormalisation [24] untested.
• Independent replication. Every result file is published with its record’s hash, but no one other than the author has yet re-run them.
• Eighteen runs without a dataset hash, completed from their own files after a watermark-band failure; their trained subsets were verified against the plan.

Download

How to cite

Green, C. (2026). Measured Influence: Pre-registered, Power-gated Validation and Calibration of Per-output Training-data Attribution for Consented Music Adapters (Working paper, version 3). Zenodo. https://doi.org/10.5281/zenodo.23618202

Contact

Questions and replication requests to khaos@khaos.dev.