Partial-Label Evaluation of Prototype Components with Temporal Aggregation

This preprint studies partial-label evaluation for interpretable prototype components in bioacoustic deployment. Using a 234-species multi-taxon soundscape testbed built from frozen Perch embeddings and a fixed ProtoSSMv2+SED rank-fusion core, the study compares covered-species offline evaluation against hidden-test full-system deployment. The results show that a prototype head can improve the covered-species offline endpoint while failing to improve the full deployment system, which demonstrates a model-selection risk when offline labels cover only part of the deployment label space. Code, result summaries, figures, and reproducibility artifacts are available in the associated GitHub repository.