What the completed V2 experiment changes
All 72 planned evaluations completed: 12 representations, three ordinary splits, and three distinct recording-date-grouped folds. The batch includes 36 newly trained representations, each trained for 50 epochs. Seven regression tests passed, and an independent verification reproduced all saved scores and checked split membership, training-only scalers, source/input hashes, and complete training histories. No probe convergence warnings were recorded. New V2 notebooks load the finished results; original notebooks were not rerun or overwritten.
The original ID-supervision headline does not survive this experiment.
Ordinary-split identity macro-F1, mean ± sample SD:
| Representation | ID macro-F1 |
|---|---|
| Simple ICI + temporal mel | 0.724 ± 0.039 |
| Supervised dual encoder | 0.684 ± 0.081 |
| DCCE, swapped channels + ID loss | 0.641 ± 0.066 |
| DCCE, whole-coda pairing without ID loss | 0.627 ± 0.057 |
| DCCE, whole-coda pairing with ID loss | 0.588 ± 0.079 |
Adding the ID objective to the whole-coda condition changes mean ID F1 by -0.039 in this fixed configuration; it does not reproduce the original large improvement. Actual channel swapping improves ordinary ID relative to the whole-coda ID condition but reduces coda-type performance, and its ID advantage does not persist in grouped evaluation. These are descriptive means over three runs, not significance claims.
The historical 0.834 is not a valid benchmark for a simple numerical before/after comparison. The revision changes split leakage, training size, ID eligibility, audio preprocessing, class weighting, and the support convention. Do not attribute the entire change to any one repair.
DCCE retains useful joint task performance, but superiority is unproven.
Without ID loss, DCCE achieves ordinary-split unit/type F1 of 0.907/0.800. It retains much more type information under this probe than the tested frozen audio embeddings. However, annotated ICI is an additional task-informed input, and the simple concatenation and supervised dual-input controls weaken an architecture-only or contrastive-objective superiority claim. Raw ICI remains particularly strong for the rhythm-defined type task.
The grouped evaluation is an important empirical result.
Social-unit macro-F1 changes from ordinary to grouped evaluation as follows:
| Representation | Ordinary mean | Date-grouped mean |
|---|---|---|
| WhAM frozen | 0.906 | 0.667 |
| AVES frozen | 0.910 | 0.700 |
| DCCE without ID loss | 0.907 | 0.687 |
| DCCE with ID loss | 0.889 | 0.574 |
DCCE is not immune to the harder split. These are different train/test partitions and do not establish that recording artifacts alone cause the decline. All unit tests contain the same three unit classes, but identity support changes substantially across grouped folds; consult the coverage tables before comparing ID scores between protocols.
Correct date parsing yields Cramér's V=0.438 over all 1,383 clean codas. Validation-only WhAM unit/year layer-score correlations are 0.263 on the first ordinary split and -0.490 on the first grouped fold. They use a new protocol and cannot simply replace the old test-profile rho as if only a typo changed. Treat them as descriptive diagnostics, not proof that identical model components encode both variables.
Recommended paper framing
Present a careful study of explicit timing inputs, acoustic representations, supervision, and evaluation conditions in low-resource sperm-whale coda classification. The evidence supports reporting channel complementarity and sensitivity to recording-date separation. It does not currently support presenting DCCE or its ID loss as broadly superior. A null or mixed result for the proposed contrastive mechanism should remain visible.
The next manuscript step is to replace the old tables, narrow claims, distinguish whole-coda pairing from channel swapping, and disclose the split/metadata corrections. Remaining experiments include a fully label-free condition, standardized single-channel ablations, bandwidth sensitivity, and stronger fine-tuning controls. The complete reviewer response map and replacement method text are in this folder.
Registry and reproducibility
The canonical folders are scripts_v2, notebooks_v2, reports_v2, and results_v2. registry_v2 snapshots preserved pre-revision material and records hashes. Compatibility symlinks retain earlier links and the exact output path embedded in the completed training source. A relocation resume check skipped all 36 already-completed neural runs; no models were retrained during organization. Checks confirmed 174 original scientific/input files unchanged, including all original notebooks. The only update to an original tracked document is the root README pointer to V2.
AVES and BirdNET caches were also spot-checked by fresh extraction for examples covering all four native sample rates plus the last clean row. Both checks passed; AVES differed only near floating-point precision and BirdNET matched exactly. This is not a full cache re-extraction or a pretraining-overlap audit. WhAM's original pretraining overlap remains unresolved.