Preserved revision-stage report. Some manuscript pointers and version labels reflect an earlier writing stage. Use the current study and research record for the final interpretation.
Final interpretation of the completed extension
All 114 main V3 evaluations, 24 explicit year-holdout evaluations, and 30 input-matched frozen-feature evaluations are complete. See year_holdout.md for all four held-out years. These findings are exploratory; neither post-hoc best-model ranking nor three-run differences establish statistical or biological superiority.
Main findings
- The historical 0.834 ID score and +0.380 improvement are withdrawn: the original encoder/probe partitions overlapped. The revised comparisons use shared holdouts and validation-only layer selection.
- Simple features are strong references: raw ICI gives 0.987 ± 0.009 random-split type macro-F1, and ICI plus temporal mel gives 0.724 ± 0.039 for ID in the corrected benchmark.
- In the uniform-sampling extension, DCCE with unit pairing and type loss gives random/grouped ID macro-F1 0.597 ± 0.051 / 0.248 ± 0.130; adding the ID objective gives 0.581 ± 0.012 / 0.223 ± 0.175. These are configuration-specific effects, not evidence for a universal benefit from an ID loss.
- Label-free representation training (with annotated timing input) gives random/grouped ID macro-F1 0.624 ± 0.064 / 0.181 ± 0.133. Unit, type, and ID objectives must be separated when interpreting the architecture.
- Input-matched pretrained controls, strict year holdouts, and channel diagnostics are reported separately in
feature_fusion.md,year_holdout.md, andbranch_probes.md. Different test support and pretraining histories limit direct generalization claims.
Paired changes
| Contrast | Protocol | Task | Mean difference ± SD | Runs |
|---|---|---|---|---|
| 16 minus 8 kHz, ID | date_grouped | coda_type | +0.042 ± 0.064 | 3 |
| 16 minus 8 kHz, ID | date_grouped | individual_id | -0.050 ± 0.047 | 3 |
| 16 minus 8 kHz, ID | date_grouped | unit | +0.016 ± 0.035 | 3 |
| 16 minus 8 kHz, ID | random | coda_type | -0.008 ± 0.017 | 3 |
| 16 minus 8 kHz, ID | random | individual_id | +0.042 ± 0.052 | 3 |
| 16 minus 8 kHz, ID | random | unit | -0.014 ± 0.018 | 3 |
| 16 minus 8 kHz, no ID | date_grouped | coda_type | +0.016 ± 0.062 | 3 |
| 16 minus 8 kHz, no ID | date_grouped | individual_id | +0.002 ± 0.002 | 3 |
| 16 minus 8 kHz, no ID | date_grouped | unit | +0.027 ± 0.080 | 3 |
| 16 minus 8 kHz, no ID | random | coda_type | +0.046 ± 0.070 | 3 |
| 16 minus 8 kHz, no ID | random | individual_id | +0.034 ± 0.045 | 3 |
| 16 minus 8 kHz, no ID | random | unit | +0.001 ± 0.014 | 3 |
| add ICI to AVES FT | date_grouped | coda_type | +0.139 ± 0.080 | 3 |
| add ICI to AVES FT | date_grouped | individual_id | +0.013 ± 0.021 | 3 |
| add ICI to AVES FT | date_grouped | unit | +0.015 ± 0.011 | 3 |
| add ICI to AVES FT | random | coda_type | +0.223 ± 0.064 | 3 |
| add ICI to AVES FT | random | individual_id | +0.130 ± 0.094 | 3 |
| add ICI to AVES FT | random | unit | -0.011 ± 0.009 | 3 |
| add ID | date_grouped | coda_type | -0.064 ± 0.149 | 3 |
| add ID | date_grouped | individual_id | -0.025 ± 0.075 | 3 |
| add ID | date_grouped | unit | -0.002 ± 0.026 | 3 |
| add ID | random | coda_type | -0.130 ± 0.040 | 3 |
| add ID | random | individual_id | -0.016 ± 0.060 | 3 |
| add ID | random | unit | +0.006 ± 0.024 | 3 |
| same-coda minus unit pairing | date_grouped | coda_type | +0.121 ± 0.003 | 3 |
| same-coda minus unit pairing | date_grouped | individual_id | -0.063 ± 0.075 | 3 |
| same-coda minus unit pairing | date_grouped | unit | +0.024 ± 0.221 | 3 |
| same-coda minus unit pairing | random | coda_type | +0.090 ± 0.057 | 3 |
| same-coda minus unit pairing | random | individual_id | +0.190 ± 0.025 | 3 |
| same-coda minus unit pairing | random | unit | -0.168 ± 0.026 | 3 |
| swap minus whole | date_grouped | coda_type | -0.157 ± 0.059 | 3 |
| swap minus whole | date_grouped | individual_id | -0.034 ± 0.107 | 3 |
| swap minus whole | date_grouped | unit | +0.025 ± 0.051 | 3 |
| swap minus whole | random | coda_type | -0.027 ± 0.022 | 3 |
| swap minus whole | random | individual_id | +0.076 ± 0.077 | 3 |
| swap minus whole | random | unit | -0.054 ± 0.004 | 3 |
Differences are paired by the same split/seed. SD is descriptive across three runs, not a significance test. The factorial changes sampling/augmentation relative to V2, so V2–V3 differences do not isolate one factor.
Within-unit identity probes
Each classifier is refitted within its social unit. Omitted scores have fewer than two supported identities in training or test. Macro-F1 is conditional on each stratum’s supported classes; do not pool these as equivalent populations.
random
| Model | Unit | ID mean ± SD | Valid runs |
|---|---|---|---|
| AVES_FT | A | 0.456 ± 0.320 | 3 |
| AVES_FT | D | 0.333 ± 0.000 | 3 |
| AVES_FT | F | 0.430 ± 0.033 | 3 |
| AVESICIFT | A | 0.722 ± 0.241 | 3 |
| AVESICIFT | D | 0.778 ± 0.385 | 3 |
| AVESICIFT | F | 0.559 ± 0.041 | 3 |
| Dual16kUTI0 | A | 1.000 ± 0.000 | 3 |
| Dual16kUTI0 | D | 0.778 ± 0.385 | 3 |
| Dual16kUTI0 | F | 0.622 ± 0.048 | 3 |
| Dual16kUTI1 | A | 0.921 ± 0.137 | 3 |
| Dual16kUTI1 | D | 1.000 ± 0.000 | 3 |
| Dual16kUTI1 | F | 0.624 ± 0.010 | 3 |
| Dual_U0T0I0 | A | 0.867 ± 0.115 | 3 |
| Dual_U0T0I0 | D | 0.556 ± 0.385 | 3 |
| Dual_U0T0I0 | F | 0.594 ± 0.026 | 3 |
| Dual_U0T0I1 | A | 0.782 ± 0.209 | 3 |
| Dual_U0T0I1 | D | 0.778 ± 0.385 | 3 |
| Dual_U0T0I1 | F | 0.588 ± 0.071 | 3 |
| Dual_U0T1I0 | A | 0.933 ± 0.115 | 3 |
| Dual_U0T1I0 | D | 0.556 ± 0.385 | 3 |
| Dual_U0T1I0 | F | 0.609 ± 0.068 | 3 |
| Dual_U0T1I1 | A | 1.000 ± 0.000 | 3 |
| Dual_U0T1I1 | D | 0.556 ± 0.385 | 3 |
| Dual_U0T1I1 | F | 0.607 ± 0.075 | 3 |
| Dual_U1T0I0 | A | 0.794 ± 0.208 | 3 |
| Dual_U1T0I0 | D | 0.556 ± 0.385 | 3 |
| Dual_U1T0I0 | F | 0.472 ± 0.072 | 3 |
| Dual_U1T0I1 | A | 0.782 ± 0.209 | 3 |
| Dual_U1T0I1 | D | 0.778 ± 0.385 | 3 |
| Dual_U1T0I1 | F | 0.445 ± 0.063 | 3 |
| Dual_U1T1I0 | A | 1.000 ± 0.000 | 3 |
| Dual_U1T1I0 | D | 0.556 ± 0.385 | 3 |
| Dual_U1T1I0 | F | 0.647 ± 0.093 | 3 |
| Dual_U1T1I1 | A | 0.921 ± 0.137 | 3 |
| Dual_U1T1I1 | D | 0.778 ± 0.385 | 3 |
| Dual_U1T1I1 | F | 0.552 ± 0.030 | 3 |
| DualswapUTI | A | 0.854 ± 0.128 | 3 |
| DualswapUTI | D | 1.000 ± 0.000 | 3 |
| DualswapUTI | F | 0.542 ± 0.042 | 3 |
| concat_UTI0 | A | 1.000 ± 0.000 | 3 |
| concat_UTI0 | D | 1.000 ± 0.000 | 3 |
| concat_UTI0 | F | 0.629 ± 0.027 | 3 |
| concat_UTI1 | A | 1.000 ± 0.000 | 3 |
| concat_UTI1 | D | 1.000 ± 0.000 | 3 |
| concat_UTI1 | F | 0.617 ± 0.043 | 3 |
| rhythm_UTI0 | A | 1.000 ± 0.000 | 3 |
| rhythm_UTI0 | D | 0.556 ± 0.385 | 3 |
| rhythm_UTI0 | F | 0.549 ± 0.046 | 3 |
| rhythm_UTI1 | A | 1.000 ± 0.000 | 3 |
| rhythm_UTI1 | D | 0.778 ± 0.385 | 3 |
| rhythm_UTI1 | F | 0.545 ± 0.048 | 3 |
| spectral_UTI0 | A | 0.514 ± 0.434 | 3 |
| spectral_UTI0 | D | 0.778 ± 0.385 | 3 |
| spectral_UTI0 | F | 0.369 ± 0.022 | 3 |
| spectral_UTI1 | A | 0.782 ± 0.209 | 3 |
| spectral_UTI1 | D | 0.778 ± 0.385 | 3 |
| spectral_UTI1 | F | 0.368 ± 0.022 | 3 |
date_grouped
| Model | Unit | ID mean ± SD | Valid runs |
|---|---|---|---|
| AVES_FT | F | 0.057 ± 0.030 | 3 |
| AVESICIFT | F | 0.078 ± 0.030 | 3 |
| Dual16kUTI0 | F | 0.283 ± 0.109 | 3 |
| Dual16kUTI1 | F | 0.186 ± 0.121 | 3 |
| Dual_U0T0I0 | F | 0.184 ± 0.130 | 3 |
| Dual_U0T0I1 | F | 0.159 ± 0.054 | 3 |
| Dual_U0T1I0 | F | 0.211 ± 0.094 | 3 |
| Dual_U0T1I1 | F | 0.165 ± 0.071 | 3 |
| Dual_U1T0I0 | F | 0.243 ± 0.087 | 3 |
| Dual_U1T0I1 | F | 0.133 ± 0.040 | 3 |
| Dual_U1T1I0 | F | 0.267 ± 0.115 | 3 |
| Dual_U1T1I1 | F | 0.238 ± 0.153 | 3 |
| DualswapUTI | F | 0.189 ± 0.070 | 3 |
| concat_UTI0 | F | 0.240 ± 0.173 | 3 |
| concat_UTI1 | F | 0.241 ± 0.046 | 3 |
| rhythm_UTI0 | F | 0.349 ± 0.138 | 3 |
| rhythm_UTI1 | F | 0.332 ± 0.142 | 3 |
| spectral_UTI0 | F | 0.054 ± 0.047 | 3 |
| spectral_UTI1 | F | 0.089 ± 0.014 | 3 |
What remains outside this pipeline
External-population validation, verified recording-session metadata, and independently excluded historical pretraining overlap are not available. No amount of rerunning these codas establishes those facts. The original numerical superiority claim is not retained. The finished paper is a reviewable research revision, not a guarantee of acceptance.