V3 exploratory extension — fixed before execution
V2 test outcomes have been inspected. This is an exploratory, prospectively specified extension, not a preregistered confirmatory study. Original materials and all V2 artifacts remain preserved.
Use the identical V2 random seeds 0–2 and date-grouped folds 0–2, training-only scalers and class weights, fixed final-epoch checkpoints, and frozen balanced linear probes. No test-driven selection. All neural conditions use uniform shuffled training rows, not unit-balanced sampling.
Eight dual-channel conditions form the full factorial of unit pairing (U), coda-type auxiliary loss (T), and individual-ID auxiliary loss (I). Without U, two independently augmented views of the same coda are paired; with U, distinct training codas from the same unit are paired. All include NT-Xent, temperature .07; auxiliary weights .5 each. Augmentation: independently mask one random 4-bin frequency band and one random 8-frame time band per view; add independent Gaussian noise of SD .01 to standardized nonzero ICI entries. Padding ICI entries remain at their training-standardized value. These are modest fixed perturbations, not tuned transformations. Annotation-derived ICI remains an input even in the label-free objective.
Controls: rhythm-only, spectrum-only, and 128-dimensional concatenation without a fusion MLP, each with UT and UTI. Both auxiliary heads attach to the available channel for single-channel controls. Concatenation uses type on rhythm and ID on spectrum; it is a jointly trained concatenation control, not independently trained late fusion. A true channel-swap UTI condition changes only pairing construction relative to dual UTI. Two additional dual UT/UTI conditions extend mel bandwidth from 8 to 16 kHz at fixed 48 kHz sampling rate, FFT, hop, mel count and duration. All train 50 epochs with AdamW(.001, weight_decay=.0001), cosine schedule and batch 64.
Additional pretrained control: fine-tune AVES transformer (frozen convolutional waveform frontend), both audio-only and audio+ICI, with supervised unit + .5 type + .5 ID objectives. Fixed 10 epochs; AdamW backbone LR 2e-5, new head LR .001, weight decay .0001; gradient clipping 1, microbatch 4 accumulated to approximately 64. Feasibility smoke test precedes full execution; any changes due to hardware failure will be recorded before test evaluation.
Report all conditions and all runs, including failed runs. Add recording-date cluster bootstrap confidence intervals conditional on fitted models, and within-unit identity probes where training and held-out support permit at least two identities. No claims of population-independent individual recognition or causal channel disentanglement. Historical pretraining overlap remains unresolved. Date is a grouping proxy, not a verified recording session.
Amendment: explicit year holdout diagnostic (before its execution)
Date-disjoint folds can share years and do not by themselves isolate year-associated cues. Add four complete leave-one-year-out tests (2005, 2008, 2009, 2010) for a fixed core: Raw ICI, ICI+temporal mel, WhAM, AVES, dual U1T1I0, and dual U1T1I1. These are chosen for the original recording-year question and the presence/absence of ID supervision, not to select the best V3 outcome. All four years and all six models will be reported (24 model/year evaluations, eight fresh neural trainings). This is an exploratory amendment after partial V3 random-split outcomes were visible; no confirmatory claim is made.
For each held-out year, split the remaining dates into train/validation with the first fold of five-fold StratifiedGroupKFold (unit labels, random_state=2000+held-out year). All data from the held-out year are test data. Assert disjoint years between training/validation and test, disjoint dates across all three parts, and shared membership across tasks. Neural seed is fixed at 0, using the same 50-epoch V3 definitions, uniform sampling, augmentation and training-only transformations. Frozen layers use validation-only selection. Primary scores retain the existing supported-class definition; full eligible scores and all support counts accompany them. Results across years are different held-out populations, not repeated-seed estimates. This diagnostic runs after the main grid releases the GPU.
Execution scheduling: the 16 frozen-feature year evaluations can run on CPU during the main GPU grid; only the eight new neural year trainings wait for GPU release. This changes scheduling, not the model set, splits, losses, or selection protocol.
Descriptive branch diagnostic (after partial main results)
For the saved dual U1T1I0 and U1T1I1 checkpoints on all three random splits and all three date-grouped folds, extract rhythm and spectral embeddings separately and fit the same training-only frozen probes. Compare their task scores with the already reported fusion scores. This is post-hoc descriptive analysis of 12 fixed checkpoints, not new encoder training, architecture selection, or replacement of the primary fused-representation metric. It tests whether auxiliary branch information is preserved at the measured fusion output without claiming biological disentanglement.
Input-matched frozen-feature controls (before their execution)
Add annotated ICI by direct concatenation to WhAM, AVES and BirdNET frozen features. Use all six existing random/date-grouped splits and all four explicit year holdouts (30 additional CPU-only evaluations). For multi-layer encoders append ICI at every candidate layer and select the layer using validation only, exactly as for the audio-only baseline. Probe scaling, C=1, class weights and eligibility remain unchanged. No encoder or auxiliary classifier is trained. This post-hoc exploratory addition was motivated by input-access fairness after partial V3 outcomes were available; all three pretrained families and all ten splits are reported, with no best-model selection or replacement of the prespecified DCCE grid.