Data for: Evidence-grounded feedback improves singer-disjoint reference-action prediction

Aggregate data companion (version 3, 2026-11-02) for the manuscript under review. Per-singer/per-seed metrics, ablation and support tables, per-action counts, rule-baseline rows and per-window table, support/query and split manifests, model and amount manifests, summary statistics, and the inventory with SHA-256 for every other file. Original VocalSet audio is not redistributed. Supersedes version 2.