Soundalike active listening study

Listen to one seed and candidates A–D. Rank all four from closest to least close in overall sound, then choose one primary reason the fourth-ranked candidate misses.

Each player is limited to an approximately 20-second strongest-recurrence excerpt. This is a repeated-section heuristic, not a verified chorus label. Candidate ordering prioritizes disagreements that are most useful to resolve.

For every vocal track, three separate 30-second positions must agree on one confidently detected singing language. Vocal candidates must exactly match the seed language; detector conflicts and unknowns are excluded. This is development/model-improvement evidence, not an independent promotion holdout, and it does not change production recommendations.