Reviewer response draft

Completed local revision: 72 corrected benchmark evaluations plus 168 extension evaluations, with all planned experiments reported and verified. This response is prepared for author review and has not been sent to anyone.

Highest priority: evaluation validity and fairness (R1 and R4)

We agree that the original comparison conflated architecture with supervision. We now distinguish unit labels used for positive pairing, type and individual-ID auxiliary labels, label-dependent sampling, annotated timing inputs, and labels used only by the downstream probe. The corrected benchmark includes a supervised dual-input encoder and AVES adapters with and without ICI. The extension adds all eight combinations of unit/type/ID training supervision, uniform sampling in every condition, and AVES transformer fine-tuning with and without ICI. Its convolutional waveform frontend remains frozen, which we disclose explicitly. These controls match downstream label availability, not pretraining histories, model capacity, or optimization budgets.

Our audit additionally found that the original separately generated identity-probe split allowed 127 of its 153 test codas into representation training. We replaced it with one shared partition per run and retrained the dataset-trained representations. We therefore withdraw the historical 0.834 ID score and the claimed +0.380 improvement. The corrected V2 whole-coda comparison gives ID macro-F1 0.627 ± 0.057 without the identity auxiliary loss and 0.588 ± 0.079 with it. It does not establish the originally claimed gain. Because multiple protocol defects were corrected together, this difference from the historical result cannot be attributed to leakage alone.

We also completed input-matched frozen-feature controls: annotated ICI is concatenated with WhAM, AVES, or BirdNET before the same training-only probe, on all six random/date partitions and four year holdouts. Their random-split ID means are 0.598, 0.652, and 0.580 respectively. This separates access to timing input from learned fusion without suggesting equivalent external pretraining. These additional controls were specified after partial extension results were visible and are explicitly exploratory.

The uniform-sampling V3 factorial independently gives random-split ID macro-F1 0.597 ± 0.051 without the ID loss and 0.581 ± 0.012 with it; grouped means are 0.248 ± 0.130 and 0.223 ± 0.175. These are descriptive three-run results, not a universal conclusion about identity supervision.

Evidence: reports_v2/revision_protocol.md, reports_v3/protocol.md, saved per-run training IDs, source hashes, shared split hashes, verification reports, paper Sections 2–4.

R1: standardized ablations and actual pairing

We now name complete-coda unit pairing and actual channel swapping separately. The old code paired f(rA,sA) and f(rB,sB), despite describing f(rA,sB) and f(rB,sA). Both implementations are evaluated explicitly. The extension includes rhythm-only, spectral-only, and jointly trained concatenation controls with and without ID loss, fixed type/ID coefficients, and identical uniform sampling. The concatenation control is 128-dimensional; fused DCCE and single-channel embeddings are 64-dimensional. We call this a concatenation control rather than implying independently trained late fusion. All probe training occurs after freezing the learned representation.

The supervision matrix in the final report enumerates U/T/I, sampling, input channels, pairing, representation training, and probe training for each condition. We no longer attribute performance to architecture merely because one row uses a different set of labels.

R1: pretrained layer selection

WhAM and AVES layers are selected solely by validation macro-F1; ties choose the earliest layer. BirdNET supplies one cached representation rather than a searched layer collection. Training-only scalers and balanced logistic regression with fixed C=1 are used throughout. Test predictions are generated after these choices. A regression test mutates test features and labels and checks that layer selection does not change. Historical test-selected values are superseded, not relabeled as unbiased results.

R1: multiple splits, intervals, and hyperparameters

We use three random seeds and three distinct recording-date-disjoint outer folds, reporting mean and sample standard deviation. We additionally resample recording dates 2,000 times within each fitted test prediction set for conditional 95% bootstrap intervals. These intervals do not capture retraining uncertainty. No interval is estimated when supported test data cover fewer than two recording dates; those cases are explicitly marked. They use the original scored class set and assign zero F1 to classes absent from a resample; this matters for rare identities. We do not treat repeated random test codas as independent observations by pooling their predictions.

Class weights are computed on training data only. The V2 benchmark explicitly uses unit-balanced sampling; the factorial extension uses uniform shuffled rows. Temperature .07, loss weights .5, optimization settings, epoch budgets, and augmentation strengths were fixed for each batch rather than selected on test results. The extension is explicitly exploratory because V2 outcomes had already been inspected. Three runs remain a limited basis for precision; we make no blanket significance claims.

R1: label-free pairing

The U0/T0/I0 condition uses two independently augmented views of each coda, with no unit/type/ID labels in either the representation objective or sampler. Other no-U conditions retain same-coda positives while adding selected auxiliary labels. All eight combinations are included. Labels still define the downstream supervised probe and evaluation split, and annotated ICI still provides one input; “label-free” refers specifically to representation training objectives and sampling.

R1: temporal feature floor and bandwidth

The mel floor now concatenates means and standard deviations from four temporal bins, calculated over valid frames only. A second simple control concatenates those statistics with ICI. It is a strong ID reference in the corrected benchmark (0.724 ± 0.039 random-split macro-F1), and raw ICI remains the strongest type reference there (0.987 ± 0.009).

The extension compares 0–8 and 0–16 kHz mel ceilings for otherwise matched dual models with and without ID loss. Sampling rate, FFT, hop, mel count, and physical time support are fixed. These tests assess the specified bandwidth change; they do not prove that information outside either band is absent.

Because the mel count is fixed, raising the ceiling also changes filter spacing below 8 kHz. We therefore do not interpret this as isolating high-frequency identity information alone.

R1: dependence of identity on social unit and coda type

The V2 report retains per-unit and per-type breakdowns of global-probe predictions, with support counts. The extension adds separately fitted within-unit ID probes, requiring at least two supported identities in both training and held-out data; one-identity strata are not reported as meaningful perfect identification. Sparse coda-type strata remain descriptive. These analyses help expose shortcuts but cannot remove every dependence between recording conditions, type, and caller.

R2: sample rates and physical preprocessing

The audio inventory found four native rates: 44.1, 48, 96, and 120 kHz. All newly computed mel features resample to 48 kHz before applying a 2,048-sample FFT and 1,152-sample hop. The revised paper specifies this transformation, sample-local log-power normalization, silence-level padding, and 256-frame support. No mel input is truncated. Pretrained encoders retain their own frontends: AVES at 16 kHz, the released WhAM codec interface at 44.1 kHz, and BirdNET at 48 kHz with a three-second window. Three of 1,383 waveforms exceed that window. This is an implemented correction, not merely an assumption that the old code already resampled correctly.

R2: loss presentation, novelty, and editing

The method now defines the auxiliary classification losses and their coefficients together with the total objective, before discussing the pairing alternatives. We removed the broad novelty claim contradicted by Tolkova et al. Their 2026 study identifies North Atlantic right whales from 234 upcalls of 11 individuals at three sites; it is relevant prior work, but not the same species or coda dataset. Our contribution is scoped to the dual-input coda study and its evaluation controls. The revised final paragraph is re-typeset and the compiled PDF is checked, including the reported missing-space issue.

R2: generalization and relevance

We retain the single-population, three-unit limitation and do not imply external biological validation. The reusable methodological findings concern explicit supervision accounting, simple annotation-derived controls, consistent multi-task holdout membership, and sensitivity to grouped evaluation. Date-disjoint testing is stronger than random splitting for the available metadata, but does not establish transfer across populations, unseen animals, or species.

R4: recording-year and session confounding

Explicit day/month/year parsing restores all 1,383 clean codas to the year analysis; the previous parser dropped 690. The full year-by-unit contingency table yields Pearson chi-squared 530.04 (df=6) and uncorrected Cramér’s V=0.438, replacing V=0.51. The year association and layerwise prediction correlations are distinct analyses, with their own sample/selection protocols. We do not preserve the old rho=.63 as though it were reproduced under the corrected validation-only procedure.

Three outer recording-date-disjoint folds now evaluate every model family. We call dates a session proxy because verified session identifiers are unavailable. Grouping substantially reduces DCCE as well as pretrained-encoder unit performance; we remove the claim that DCCE is generally less susceptible. ID class coverage changes across grouped folds and is reported explicitly. Score declines alone do not identify causal recording artifacts.

R4: bibliography

The author list for Andreas et al. is rebuilt from publisher-deposited DOI metadata, correcting the omitted and incorrect names in the original reference [1]. We also resolve source-version discrepancies in the WhAM and Sharma references, and use the published 2026 Tolkova record. The final bibliography is unified in latex_v3/refs.bib; raw DOI metadata is preserved in registry_v3/bibliography_metadata.json.

Remaining scientific limitations

No independent whale population or fully untouched external dataset is available in this workspace. Historical pretraining overlap cannot be independently excluded. Recording dates are not verified sessions. The exploratory extension does not become confirmatory merely because its settings were fixed before its own execution. These limitations are disclosed; they are not replaced by assertions that more local reruns resolve them.

In particular, WhAM's published training uses DSWP and CETI, while our evaluation collection also comes from DSWP. Exact file membership for the released checkpoint remains unverified. Clean local holdouts and numerical cache checks do not establish independence from that earlier pretraining.

Additional explicit year diagnostic

All 24 core model/year evaluations are complete and verified: four strict year holdouts for Raw ICI, ICI+temporal mel, WhAM, AVES, and dual U1T1I0/U1T1I1. This removes shared years between local train/validation and test, which date-disjoint folds alone do not guarantee. All four years and six models are reported, with supported and full-eligible scores and coverage counts; the three input-matched frozen-feature controls add 12 further year evaluations. The amendment was made after partial random-split V3 results had been inspected and remains exploratory. DCCE 110 unit macro-F1 is 0.592/0.362/0.767/0.346 for 2005/2008/2009/2010; DCCE 111 gives 0.448/0.391/0.713/0.357. These outcomes do not support general resistance to recording-year variation. Full task results are in reports_v3/year_holdout.md and reports_v3/feature_fusion.md.

Diagnostic branch probes and cache checks

All 12 saved DCCE 110/111 checkpoints were reloaded; their reconstructed fusion embeddings exactly match the saved arrays. Separate rhythm/spectral probes are post-hoc diagnostics, not replacements chosen for the primary scores. Adding ID supervision raises random-split spectral-branch ID from 0.367 to 0.443 while fused ID changes from 0.597 to 0.581. Rhythm branches retain type scores of 0.983/0.982. Branch-level auxiliary performance therefore need not carry through fusion, and these results do not establish biological disentanglement. The full-coda spectrogram itself also contains timing information.

Five sampled re-extractions per pretrained encoder agree with the cached embeddings. WhAM checks span all four native sample rates and have relative L2 error below 2.4e-7. These checks establish sampled numerical consistency, not historical pretraining independence.