Reviewer revision audit · DCCE research

Historical report; not the current benchmark. This document retains original claims and scores, some of which were corrected or withdrawn. Use the current study and research record for the final interpretation.

Review priorities, manuscript comparison, and correctness audit — 20 September 2026

The reviews are R1: weak accept, R2: weak accept, R4: weak reject. All see merit in the biological motivation. R4's rejection is principally about unequal supervision and missing evaluation that controls recording conditions. These are the highest-priority reviewer concerns. However, inspecting the implementation revealed additional correctness problems that must be fixed before interpreting new or existing results.

This report covers all actionable reviewer comments, consolidating repeated comments. It compares final_paper.md, latex/main.tex, latex/main_beyondwham_v1.tex, and extracted text from latex/main.pdf; checks training/evaluation scripts and notebook source; recomputes dataset and split diagnostics; and spot-checks relevant primary literature. It does not rerun model training or certify the historical provenance of every published score. Findings describe the files currently in this workspace. The actual submitted paper PDF was not supplied, so an exact submitted-version comparison remains unavailable.

Recommended priority order

Effort below means relative implementation and analysis effort, not a promise of elapsed time. P0 means fix before relying on results; P1 means central to answering reviewers; P2 means supporting experiments; P3 means editorial work. Cheap P3 edits can proceed alongside experiments.

Rank	Priority / origin	Work required	Effort	Completion evidence
1	P0 — audit	Repair encoder/probe split leakage, label eligibility, and train-only preprocessing.	Medium + full retraining	Shared split manifest; no test coda used in representation training; corrected ID results.
2	P0 — audit	Reconcile claimed cross-channel swapping with the actual objective.	Medium–high	Explicit implementation and equations agree; rerun results for the chosen method.
3	P0 — audit; R4 confound	Fix date parsing and regenerate all year analyses.	Low for counts; medium for probes	All 1,383 dates parse; corrected contingency, year probes, plots, and claims.
4	P0 — R1 best-layer question	Select baseline layers and all tuned settings on validation data only.	Medium	Logged validation selections and untouched test evaluation.
5	P1 — R1/R4 supervision fairness	Add comparisons with equivalent access to unit, type, and ID labels.	High	Explicit supervision matrix and matched supervised controls.
6	P1 — R4 recording confound	Evaluate every principal model on year-disjoint and/or recording-session-disjoint splits.	High	Group-disjoint manifests, class coverage, paired results and uncertainty.
7	P1 — R1 ablations	Standardize supervision and loss weights across architecture ablations; test with/without ID.	Medium–high	Ablation matrix that isolates one factor at a time.
8	P1 — R1 variability/tuning	Repeat complete training/evaluation over predeclared splits/seeds; explain all tuning.	High compute	Per-run scores, mean/SD, appropriate uncertainty, paired differences.
9	P1 — R2 sample rates; audit	Standardize physical time/frequency preprocessing and regenerate features.	Medium + retraining	Verified audio-rate inventory; explicit resampling/STFT/padding configuration.
10	P2 — R1 label-free pairing	Compare same-coda augmented pairing with unit-label pairing.	Medium	Separate pairing-label and fully label-free controls, with other factors matched.
11	P2 — R1 ID strata	Report identity performance within social unit and by coda type.	Low–medium after reruns	Support counts, candidate classes, per-stratum scores, uncertainty.
12	P2 — R1 stronger mel floor	Add temporal mel pooling or a simple 2D-mel model and a concatenated ICI+mel baseline.	Medium	Same splits, comparable supervision, validation-selected settings.
13	P2 — R1 spectral bandwidth	Compare 0–8 kHz with 0–16 kHz and, if useful, a high-band-only control.	Medium + retraining	Fixed physical time support and comparable model capacity across bands.
14	P1 writing — R2/R4 scope/impact	Narrow claims, clarify species-specific novelty, and articulate reusable methodological lessons.	Low–medium	Revised introduction/discussion tied to corrected evidence.
15	P3 — R4 references; audit	Verify all bibliographic records, including reference [1], WhAM, and Sharma.	Low	One authoritative bibliography and checked author lists.
16	P3 — R2 organization/typo; version audit	Define losses where introduced; fix notation/spacing; consolidate conference source and build assets.	Low–medium	One compiling conference manuscript with synchronized figures and tables.

Correctness issues found beyond the reviews

Individual-ID evaluation is not held out from encoder training. The saved general split contains 1,106 training and 277 test codas. The separately generated ID split contains 609 training and 153 test codas. Mapping both to coda IDs shows 127/153 ID test codas overlap encoder training; 129 ID training codas fall in the general test split. scripts/lib_eval.py uses these independent splits. scripts/dcce_fairness.py::train_full trains on the general split and supplies identity labels for eligible codas, without excluding the ID test set. The Phase 3 builder/notebook follow the same pattern. For the with-ID variant this exposes evaluation identities for overlapping samples directly to the training loss. For the no-ID variant, overlapping samples still supply audio, type, and unit information. The advertised ID scores, including 0.834 and 0.571, cannot be treated as clean inductive held-out estimates under this pipeline. The saved split overlap is proven; historical checkpoints still need run-level provenance verification.

Fix: assign train/validation/test membership once at coda or recording-group level, then derive every task subset from that manifest. Restrict all representation-training views, partner sampling, label losses, and learned preprocessing to training data. Preserve class coverage where possible and state when a task cannot be evaluated. Retrain DCCE and any adapted baselines. Cached frozen external embeddings can generally be reused if their extraction was sample-local; pretraining-data overlap still needs disclosure.
The central pairing operation differs from the paper. final_paper.md:153 describes f(r_A, s_B) paired with f(r_B, s_A). Both scripts/dcce_fairness.py::train_full and builders/build_phase3_notebook.py:568 actually compute f(r_A, s_A) paired with f(r_B, s_B). The stored Phase 3 notebook source agrees with the code. The implemented experiment is same-unit contrastive pairing of complete codas, not channel-swapped fusion. Choose whether to describe the implemented method accurately or implement the proposed swaps and rerun. Do not simply retain the equations and reuse the old scores. An explicit swapped-versus-unswapped control would test the proposed novelty.

Additional objective details: NT-Xent treats other same-unit samples as negatives as well as other-unit samples; the paper's description of negatives spanning other units is incomplete. WeightedRandomSampler balances unit counts in expectation, not exactly within each batch. Partner selection does not exclude the anchor itself. The late-fusion control duplicates z with a detached clone; it is not a separately augmented pair. These are design choices requiring accurate reporting and, where changed, new runs.
Date parsing silently removes half the dataset from the confound analysis. builders/build_phase2_notebook.py:80 uses pd.to_datetime(..., errors="coerce") without the day/month/year format. In the current environment, this makes 690/1,383 dates missing. The remaining 693 codas reproduce the stored notebook table and the manuscript's chi-squared 361.67 and V=0.5108. Explicit %d/%m/%Y parsing retains all 1,383 records and gives chi-squared 530.0436, df=6, p≈2.83e-111, and Cramér's V=0.43775.

Corrected counts from the current label CSV:

Year Unit A Unit D Unit F

2005 10 0 358

2008 27 171 65

2009 30 18 47

2010 174 132 351

Consequently, “A only in 2005/2009” and “D only in 2008/2010” are false for the full clean dataset. Regenerate the recording-year probe, its across-layer correlation, and all dependent figures; do not carry over rho=0.63 without recomputation. Association still motivates a confound check, but neither the contingency test nor correlated layer scores proves that a model uses recording artifacts. Remove claims that DCCE is protected by its architecture unless controlled evaluation demonstrates it.
WhAM/AVES layers are selected on the test set. scripts/run_baselines.py::best_layer_eval maximizes scores returned by probe, whose evaluation indices are test indices. It then reports that same score and repeats splits at the selected layer. R1's concern is confirmed in the code. Multiple seeds alone do not fix test-selected layers. Use inner validation for each outer split, including probe regularization and any tuned temperature, sampling, weights, or epoch count. BirdNET is a flat embedding in this script; no layer search is performed for it.
Native audio rates are more varied than reported. Reading all 1,501 WAV headers found 394 at 44.1 kHz, 846 at 48 kHz, 42 at 96 kHz, and 219 at 120 kHz. Phase 3 preprocessing uses librosa.load(..., sr=None), so the source retains native rates. The paper's 44.1/48 kHz inventory is incomplete. With fixed 2,048/512 sample STFT sizes, a frame and the 128-frame crop cover different physical durations at different rates. At 120 kHz, the nominal 128-hop span is about 0.546 s versus 1.486 s at 44.1 kHz. This can materially change how much of a coda is retained.

Fix: audit the cached mel array's provenance, select a common sample rate adequate for the planned upper frequency band, explicitly set all STFT parameters, regenerate features, and retrain. Alternatively use time-defined windows/hops and duration handling that genuinely match across rates. Inspect zero padding: after conversion to dB relative to the peak, zero is not a silence floor; document or replace padding accordingly. R2's suggestion that resampling probably already occurs is not supported by the inspected code.
Dataset exclusions and identity labels need correction. The current CSV has 1,383 clean codas, 620 unknown-ID codas, and 763 with nonzero ID strings. A singleton (6059, coda 534) is excluded by the evaluation loader, leaving 762. Thus the paper's “621 lacked identification” conflates 620 unknowns with one identified singleton. There are 12 retained ID strings, but one is 6070/6068, covering eight codas. Resolve this combined/ambiguous identity label with the source metadata before describing the task as exactly 12 distinct identified whales. Do not arbitrarily assign those codas to one whale. The old LaTeX dataset table's ID-labeled unit counts sum to 762 but do not reflect the singleton exclusion correctly.
Ablations do not isolate only channel presence. The single-channel branches alias the same embedding as both rhythm and spectrum, then apply both type and ID heads. Their type-loss coefficient is 1, while fused variants use 0.5. Therefore rhythm-only also gets ID supervision and spectral-only also gets type supervision, not only each channel's nominal task. Make the actual objective matrix explicit and match loss coefficients, labels, training budgets, and split membership. Fused no-ID versus with-ID isolates the added ID term only after the split is repaired; it does not establish an architecture-only advantage over frozen encoders.
Other method text is out of sync. The fusion implementation includes Linear(64,64) after ReLU; the conference source omits this final layer and places L2 normalization in the architecture, whereas the code normalizes inside the contrastive loss. The type cross-entropy calls have no class-weight argument despite the paper describing weighted type loss. The probe harness uses max_iter=2000, while the conference paper says 500. The loader fits ICI StandardScaler on all clean codas; fix this for representation training. For a raw-ICI probe later standardized again on training rows, the first affine scaling need not change the result, but that does not excuse global scaling before neural training. The asserted advantage of pre-projection features over fused z needs explicit experimental support or removal.

Year	Unit A	Unit D	Unit F
2005	10	0	358
2008	27	171	65
2009	30	18	47
2010	174	132	351

How to answer each reviewer concern with adequate evidence

Fairness and supervision (R1 technical comments/general Q2–Q3; R4 principal concern). The “no ID supervision” label does not mean label-free: DCCE still receives unit labels for pair construction and type labels for an auxiliary classifier. Add separate columns for unit/type/ID labels, annotated ICI input, and pretraining source. Keep frozen pretrained models as transfer references. Add a simple model trained on the same ICI+mel inputs with the same label budget, and adapted pretrained baselines (fine-tuning or clearly identified supervised adapters). Adapters are a useful practical control but do not fully equalize encoder training. State the distinction. A same-coda pairing variant removes unit labels from pairing only; a fully label-free representation control must also remove auxiliary label losses and label-dependent sampling. Annotated ICI access must still be disclosed.

Confound-controlled evaluation (R4). First use corrected dates and inspect per-year/per-session class and identity support. Prefer actual recording-session identifiers; recording date is an available grouping proxy (44 distinct dates), not a verified session ID. Evaluate a grouped split alongside the ordinary split for every principal model. Leave-one-year-out is useful where class coverage permits, but 2005 has no Unit D and some identities may be absent in a training or test fold. Report exclusions, support, and the evaluated class set. Closed-set ID classification cannot assess an unseen whale class without changing the task; a retrieval/verification experiment is separate. A year-only unit predictor and year probes for all models would add useful diagnostics. No grouped split by itself proves biological causality or cross-population generalization.

Uncertainty and tuning (R1 validation comments/general Q1/Q5). Predeclare a feasible repeated-split design (for example five valid outer splits, with more runs if affordable), retrain every dataset-trained representation per split, and compare models on the same outer splits. Report mean/SD and paired score differences. Confidence intervals should respect recording-group dependence; bootstrap groups where feasible and explain sparse-class limitations. Do not call ±SD a confidence interval or treat overlapping random splits as independent observations. dcce_fairness.py currently also probes one trained encoder across reshuffled splits; that path must be replaced with per-split retraining. Report whether each hyperparameter was fixed in advance or selected using validation data. The already inspected test set should be acknowledged as exploratory; a new, preregistered holdout is preferable where feasible.

Identity by type/unit (R1). Report within-unit candidate pools, per-type ID results for sufficiently supported types, class counts, and a per-individual breakdown. Distinguish subsetting predictions from retraining a classifier within each unit; these answer different questions. Avoid drawing conclusions from tiny cells. The current raw ICI ID score of 0.493 also argues against claiming identity resides exclusively in spectrum.

Feature floors and bandwidth (R1 general Q4 and validation comments). Add time-aware pooling (e.g. fixed temporal bins) or a small 2D CNN on mel, and a simple concatenation of ICI with spectral features. The latter tests whether learned fusion beats merely retaining both inputs. Compare frequency bands only after sample-rate/time-support fixes; keep labels, splits, capacity, and optimization comparable. Choose settings on validation data, then report held-out outcomes, including null results.

Generalization and relevance (R2). Explicitly limit empirical conclusions to the Dominica corpus and tested splits. An external population/species experiment would strengthen scope but is not a prerequisite for an honest narrow claim; YouTube augmentation alone would not establish population generalization. Reusable contributions can be a verified evaluation protocol for confounded bioacoustic data, a test of explicit rhythm inputs alongside acoustic embeddings, and the measured effect of supervision on transfer. Present these as evidence from this case study. Avoid “what whales are saying” or semantic-decoding conclusions from classification results.

Novelty contradiction (R2). Narrow the claim to the exact sperm-whale-coda setting instead of asserting that multi-individual cetacean representations have never been evaluated. Tolkova et al. study North Atlantic right-whale upcalls: 234 samples, 11 individuals, three sites. This is relevant related work, not the same population or call type. Primary paper.

Editorial issues (R2/R4). Define type and ID losses immediately after the total-loss equation; reserve the next subsection for pair construction and contrastive loss. The reported missing space in “accuracy.By” is already correct in the supplied final_paper.md, so check the eventual PDF rather than blindly applying an edit. Check references against records, not against another local draft. Replace categorical claims about channel independence, tokenization losing timing, and DCCE being immune to drift with narrower interpretations: poor linear decoding does not prove absent information; a mel spectrogram retains time structure; architectural branches do not establish statistical or biological independence.

Which manuscript is current?

final_paper.md is actually complete LaTeX source, despite its extension. Its title and organization match the review closely, making it the best candidate conference-source baseline. This is a content-based inference, not proof of the exact uploaded submission. latex/main.tex is an expanded course/report version with the same principal results but a different presentation. latex/main_beyondwham_v1.tex is an older experimental narrative. No version is currently verified correct against the implementation.

Element	`final_paper.md`	`latex/main.tex`	Recommended treatment
Format/title	DCASE/IEEE; “Dual-Channel Contrastive Encoding of Sperm Whale Coda”	Rho report; “Two Channels, One Representation…”	Use conference format for revision.
Authors	João Victor Quintanilha, Emma Virnelli, Shuo Zhang	João Quintanilha, Emma Virnelli	Preserve intended conference author list; do not copy the report header over it.
Main/ablation scores	Seven main rows and five ablation rows	Same numerical values, including no-ID DCCE and AVES/BirdNET	Numerically synchronized, but require protocol repairs and reruns.
Supervision terminology	“no ID sup.” with remaining overclaims of fairness	Repeated “label-free” and architecture-only interpretations	Rewrite both around the complete supervision matrix.
Background	Expanded identity/type/unit discussion and speaker-recognition literature	Broader background, data-assembly account, additional references	Preserve conference improvements; transfer useful detail selectively.
Data methods	Short paragraph	Dedicated label-assembly section and dataset table	Restore reproducibility details after correcting counts.
Hyperparameters	Omits explicit epoch count	States 50 epochs and more sampling details	Report actual reproducible configuration.
Figures	Architecture PNG, two-channel plot, year plot	More figures and embedded architecture/pair diagrams	Regenerate impacted figures from corrected runs.
Uncertainty limitation	Single split, little explicit explanation	Explicit per-fold-retraining limitation	Keep explanation and replace with actual repeated training where possible.
References	Inline bibliography, including Andreas and 2026 Beguš	External `rho.bib`, different keys/content	Consolidate into one verified bibliography.
Conclusion	Explicit ID/type tradeoff and future work	More categorical channel-independence framing	Prefer qualified conference framing, revised after reruns.

The older main_beyondwham_v1.tex lacks the newer AVES/BirdNET/no-ID comparison and reports WhAM ID=0.454 plus the “+83%” headline. It retains a synthetic-augmentation experiment absent from both newer narratives. Treat it as an archive of additional experiments, not the authoritative manuscript; its results also require the evaluation audit before reuse.

The extracted latex/main.pdf text has the “Two Channels” title and two-author report header. It is not a compiled copy of the supplied conference source. Its contents broadly correspond to latex/main.tex; no byte-for-byte build equivalence or visual-layout certification was performed.

The conference source is not a self-contained build in this checkout: its figures/dcce architecture.png reference does not exist (an underscore-named latex/figures/dcce_architecture.png does), the other figure paths resolve under latex/ rather than the repository root, and no dcase2026.sty is present in the repository search. Check whether the existing architecture image actually depicts the chosen implementation before reusing it. Rename/copy the selected conference source to an unambiguous .tex file, establish its working directory/assets/style dependencies, and compile/render the final revision. Do not overwrite main.tex wholesale and lose author/content changes.

Bibliography spot checks

R4 is correct about conference reference [1]. The Andreas author list wrongly includes P. L. Tyack and R. Winn, omits N. Pavlov, P. Tønnesen, A. Torralba, and D. Vogt, and gives Wood's initials incorrectly. Replace it with the original record: Jacob Andreas; Gašper Beguš; Michael M. Bronstein; Roee Diamant; Denley Delaney; Shane Gero; Shafi Goldwasser; David F. Gruber; Sarah de Haas; Peter Malkin; Nikolay Pavlov; Roger Payne; Giovanni Petri; Daniela Rus; Pratyusha Sharma; Dan Tchernov; Pernille Tønnesen; Antonio Torralba; Daniel Vogt; Robert J. Wood. Original article.
latex/rho.bib has an incorrect Sharma author list. The conference inline entry agrees with the original list: Sharma, Gero, Payne, Gruber, Rus, Torralba, Andreas. Original article.
latex/rho.bib has incorrect WhAM names and an extra author; the conference entry's initials agree with the arXiv record. Correct full names include Pranav Muralikrishnan, Liangyuan Chen, and Hugo Flores García; Pratyusha Sharma is not on that record's author list. WhAM record.
These are targeted checks, not a certification of every citation. Verify the remaining references and the claims they support when consolidating the bibliography.

Execution order and minimum credible revision

Freeze copies of historical results and mark them provisional. Establish the conference source and experiment provenance. Resolve ambiguous identity labels and fix date/sample-rate processing.
Create a common split manifest with group-aware alternatives, training-only preprocessing, and inner validation. Add checks for coda/group overlap, label eligibility, missing dates, and class support.
Reconcile the method implementation and equations. Run a small end-to-end check before the full experiment grid.
Rerun the feature floors, frozen baselines with validation-selected layers, DCCE with/without ID, a matched supervised dual-input control, and a comparable supervised adaptation of a strong audio baseline. Evaluate ordinary and grouped splits; repeat training per outer split.
Add controlled architecture/pairing ablations, ID strata, temporal-feature floors, and bandwidth sensitivity in that order as resources permit. A pairing-only label-free condition and a fully label-free condition answer different questions.
Regenerate tables/figures from saved metrics. Rewrite the fairness, confound, novelty, and generalization claims to fit the outcomes. Finish citations and presentation. Prepare a response table linking each reviewer comment to the revised section, experiment, or justified limitation.

If resources are tight, prioritize corrected held-out evaluation, matched supervision, and grouped robustness over extra model breadth or synthetic augmentation. If these experiments do not support DCCE superiority, a defensible revision can instead center on the corrected empirical comparison and evaluation confounds; acceptance cannot be inferred from the current vote balance.

Audit artifacts and limits

Reproduce the data checks with .venv/bin/python scripts/audit_paper_review.py > reports/review_audit_evidence.json.
The JSON records input hashes, split overlap coda IDs, counts, both date tables, corrected statistics, and audio-rate inventories for all/clean data.
No paper source, experiment script, dataset, cached features, or existing result was modified for this assessment. Only this report, the audit script, and its JSON output were added.
No new predictive performance, confidence interval, architecture advantage, or clean ID score is claimed. Historical result provenance, a complete bibliography review, corrected training runs, and final conference PDF build/visual checks remain revision work.

Back to the study Explore the research archive →