The research question stayed; the conclusion changed
The original goal was to ask whether explicitly encoding rhythm and spectral information improves representations for social unit, coda type, and individual identity. That question remains central. The revised study adds stronger input and supervision controls and evaluates across recording dates and years.
The original interpretation emphasized representation superiority. The corrected evidence supports a narrower conclusion: combining inputs can be useful, but effects depend on task, supervision, available annotations, and evaluation conditions. Simple timing and combined-feature references remain strong.
Why the earlier identity result was withdrawn
The historical individual-ID score of 0.834 and claimed +0.380 improvement are not valid corrected benchmark results. The original encoder-training and identity-test partitions overlapped. The revision uses shared coda holdouts throughout local representation training and probing.
Several other elements also changed, including model-selection procedures, preprocessing, sampling, and controls. The difference between historical and corrected scores therefore cannot be attributed entirely to one leakage correction. The earlier outputs are preserved so the work can be inspected, not presented as current evidence of superiority.
What changed after review
| Concern | Response in the corrected study | Remaining limit |
|---|---|---|
| Unequal supervision | Explicit U/T/I definitions; all eight combinations; matched channel controls; supervised AVES adaptation | Pretraining, capacity, objectives, and budgets still differ |
| Unequal input information | Frozen WhAM, AVES, and BirdNET each evaluated with appended ICI | Annotated timing remains additional information |
| Ambiguous ablations | Separate whole-coda pairing, actual swapping, architecture controls, and saved-branch probes | Concatenation changes representation dimensionality |
| Test-based layer selection | WhAM/AVES layers selected per task on validation only | Study remains exploratory |
| One split | Three random splits and three date-grouped folds with sample SD | Three runs are limited; SD is not a confidence interval |
| Recording-year confound | Correct date parsing, contingency analysis, grouped evaluation, four year holdouts | Context is not experimentally isolated; sessions unverified |
| Reproducibility and traceability | Saved code, predictions, histories, source hashes, notebook outputs, and numerical reconciliation | No fresh end-to-end training replay during the final audit |
Stages of the work
The original investigation includes label assembly, EDA, simple baselines, WhAM probing, dual-encoder experiments, synthetic augmentation, and later broader pretrained baselines. These are accessible through the executed notebooks and report archive.
The first correction stage rebuilt shared splits, training-only transformations, and validation-only selection. The extended stage added matched uniform-sampling supervision controls, AVES adaptation, strict year holdouts, frozen-plus-ICI controls, and branch diagnostics. Version labels such as V2 and V3 identify these artifact families; they are not competing methods in the paper.
The artifact names preserve this sequence. They do not imply that later tests were preregistered before earlier outcomes were inspected.
Synthetic-data experiments belong to the historical research record. They are not part of the corrected main benchmark and do not establish an improvement claimed by the final paper. Likewise, early t-SNE or other embedding visualizations remain descriptive explorations, not proof of class separability or biological disentanglement.
Preserving the original execution record
All ten main notebooks retain their saved execution counts, text, tables, images, and errors. Browser-readable HTML versions let readers inspect them without executing code. The original phase3_dcce_plus_yt notebook contains one saved error and one unexecuted code cell; that state was preserved rather than silently repaired.
The newer results notebooks read completed artifacts. They do not mean that model training happened inside those notebook cells: training histories, checkpoints, result files, and logs record the underlying runs separately. Frozen baselines reuse pretrained embedding caches and evaluate probes; their foundation models were not retrained locally.
From the full study to the short paper
The conference manuscript retains the original research question and core structure while fitting four content pages plus references. Its limited space prioritizes the central comparisons, selected ablations, and necessary qualifications. This companion restores descriptive EDA, extended controls, methodological detail, and evidence links without changing the reported experiment values.
The website adapts and retains substantial manuscript wording in its methods, evaluation, and interpretation chapters. New prose organizes the larger record for reading; new descriptive figures come from existing labels and audio. No model training or probe fitting was performed to build the site.
Explore: executed notebooks, report library, numerical audit, and final paper.