Reviewer response map for the corrected first-round experiment
This is an internal revision tracker, not a claim that every concern has been fully resolved. Final numeric statements should be taken from revision_results.md after its completion counter reaches 72/72.
| Concern | New evidence or implementation | Status / remaining limitation |
|---|---|---|
| R1/R4: unequal supervision | Explicit unit/type/ID/sampling matrix; DCCE with/without ID; supervised dual-input encoder; AVES adapters with/without annotated ICI; same train/validation/test membership | Meaningful matched-label controls implemented. Adapters are not full pretrained-backbone fine-tuning; describe this accurately. |
| R1: ambiguous ablations | Whole-coda pairing and actual channel-swapped pairing separately named and trained under the same auxiliary loss weights; full training configurations saved | Central pairing and ID comparisons covered. Full single-channel and label-budget factorial still pending. |
| R1: best-layer test tuning | WhAM and AVES layer selection exclusively on validation; test labels/features mutation regression test; all validation trials saved | Addressed in corrected pipeline. Historical test-selected scores remain exploratory. |
| R1: one split/no variability | Three complete ordinary-split retrainings; three distinct grouped outer-fold retrainings; mean and sample SD; paired differences | Initial repeated evaluation implemented. SD is not a confidence interval; three folds do not justify strong significance claims. |
| R1: class weights/sampling/temperature selection | Parameters fixed before the corrected batch; train-only weights and normalization; no test-dependent epoch choice | Configuration documented. No claim that fixed parameters are optimal. |
| R1: ID dependence on type/unit | Per-unit/type/individual breakdowns saved with support counts, computed from saved global predictions | Descriptive analysis supplied. Separate within-unit classifiers and larger supported strata remain possible extensions. |
| R1: weak mel floor | Four-bin temporal mel mean/SD, plus a simple ICI+mel concatenation linear probe | Addressed with a stronger transparent floor and a dual-input control. |
| R1: label-free pairing | Correct terminology and all sources of label use documented | A genuinely label-free objective/sampling condition is not in this first grid. Do not label no-ID DCCE self-supervised. |
| R1: 8 versus 16 kHz bandwidth | Common-rate preprocessing supports a later controlled band comparison | Sensitivity experiment deferred; 0–8 kHz is a fixed choice, not proven optimal. |
| R2: sample-rate ambiguity | Inventory of four native rates; explicit 48 kHz resampling, fixed physical hop/window, silence-level padding, no truncated clips | Corrected and documented; old cached mel is not reused. |
| R2: losses introduced out of order | Replacement method text defines supervised terms with the total objective | Draft supplied; integrate into the conference manuscript. |
| R2: broad novelty conflicts with Tolkova | Scope novelty to sperm-whale codas; acknowledge multi-individual right-whale work | Editorial change identified in review audit; do not retain the broad “no learned representation” claim. |
| R2: narrow generalization and DCASE impact | Ordinary-versus-date-disjoint evidence, explicit low-resource input/supervision comparisons, public split/provenance artifacts | Strengthens methodological relevance; no external-population validation performed. |
| R2: missing final-sentence space | Supplied source already contains the space | Check compiled conference PDF; no blind duplicate edit required. |
| R4: year/session confound not controlled | Three recording-date-disjoint folds, corrected complete date table, validation-only year profiles | Date grouping is a session proxy. No claim of causal artifact attribution or perfect biological isolation. |
| R4: incorrect reference [1] | Verified author list in reviewer_revision_audit.md; additional WhAM/Sharma discrepancies found |
Bibliographic corrections identified; final unified bibliography and conference PDF remain editorial work. |
| Audit: ID test exposure during encoder training | Single shared membership; train-only partner pools and ID losses; persisted training coda IDs checked against holdout | Corrected and independently verified per run. |
| Audit: claim/code pairing mismatch | Explicit whole-coda reference plus actual swapped-channel condition | Corrected experiment design; old scores cannot be attributed to the swapped condition. |
| Audit: date parser drops 690 codas | Explicit format with errors raised; all 1,383 parsed; corrected chi-squared/V and year profiles | Corrected; old year-coverage claims and rho must be removed. |
| Audit: identity ambiguity and singleton | Original metadata checked; ambiguous/singleton labels excluded only from ID objective/evaluation | Conservative policy documented; external clarification could permit later inclusion. |
Suggested response wording for the major experimental issues:
“We agree that the original comparison did not isolate architecture from supervision. We now enumerate unit, type, identity, and sampling supervision explicitly and include a supervised dual-input control and supervised adapters on frozen AVES features, both with and without annotated timing input. All representations are evaluated with training-only probes under the same task-consistent split manifests. The adapters match downstream label access but do not constitute full encoder fine-tuning, which we state as a limitation.”
“Our audit also identified a mismatch between the representation-training split and the separately generated identity-probe split. We replaced these with one shared train/validation/test manifest per run and retrained the dataset-trained representations. We therefore replace the historical identity scores rather than presenting them as valid held-out estimates.”
“We corrected explicit day/month/year parsing, which restores 690 codas omitted from the original recording-year contingency analysis. Across all 1,383 clean codas, the corrected association is chi-squared = 530.04 with six degrees of freedom and Cramér’s V = 0.438. We additionally evaluate recording-date-disjoint folds. We describe date as a session proxy and avoid attributing score differences solely to recording artifacts.”
“The previous implementation paired complete codas from the same unit; it did not implement the channel-swapping equations in the original manuscript. We now distinguish that reference objective from an explicitly implemented swapped-channel variant and report them separately. Our revised interpretation follows the measured comparison rather than assuming that channel swapping improves the representation.”
Do not submit these responses without incorporating the corresponding manuscript changes and checking the completed results. Do not replace unresolved experiments with assertions that they succeeded.