Pretraining exposure and cache provenance

The WhAM paper, Appendix E.2 states that species-specific training uses DSWP+CETI. The local WhAM README likewise names DSWP for species-specific fine-tuning. Our evaluation collection is also drawn from DSWP. This establishes a shared corpus source, not exact sample-level overlap for the particular released checkpoint. No complete per-example checkpoint training manifest has been established in this workspace.

Consequently, random, date-disjoint, and year-disjoint partitions control the local representation/probe training implemented in this revision. They do not retroactively remove examples or years from WhAM's earlier training. WhAM remains a useful available-checkpoint reference, but its scores cannot be claimed as independently verified pretraining-disjoint performance. AVES/BirdNET have different external pretraining histories; exact historical overlap has not been independently excluded for them either.

The fresh extraction spot checks answer a different question: whether current audio and extraction reproduce archived feature vectors. They do not establish pretraining independence. V2 includes five rate-spanning checks each for AVES and BirdNET. The corresponding five-file WhAM CPU check also passed, with relative L2 errors from 1.61e-7 to 2.33e-7. The sampled files span all four native rates. Checkpoint hashes, extraction-source hash, codec rate and the LAC dependency commit are saved in results_v3/wham_cache_spotcheck.json. This is a spot check, not a full cache rebuild.

The AVES fine-tuning controls update the feature projection and all twelve transformer blocks on local training rows, while freezing the convolutional waveform frontend. They match access to downstream unit/type/ID labels, not pretraining data, capacity, optimization budget, or objective form.