Executed revision notebook: saved outputs; no cells run by this page. Read the research record · All notebooks

Complete review revision: V3

This notebook reads the preserved experiment outputs. It does not rerun training by default. V3 is exploratory because V2 results were examined before its design. See reports_v3/protocol.md for the fixed experiment grid.

from pathlib import Path
import json
import pandas as pd
pd.set_option('display.max_rows', 120)
from IPython.display import display, Markdown, Image
ROOT = Path.cwd()
if not (ROOT / 'results_v3').exists():
    ROOT = ROOT.parent
assert (ROOT / 'results_v3').exists()
OUT = ROOT / 'results_v3'
print(json.loads((OUT / 'completion.json').read_text()))

{'completed_v2': 72, 'completed_v3': 114, 'expected_v3': 114, 'completed_v3_main': 114, 'completed_year_holdout': 24, 'completed_new_total': 168, 'expected_new_total': 168, 'completed_feature_fusion': 30, 'primary_evaluations_including_v2': 240, 'diagnostic_branch_probes': 24}

display(Markdown((ROOT / 'reports_v3/supervision_matrix.md').read_text()))

V3 supervision and implementation matrix

U/T/I refer to labels used during representation training. Every row later freezes the representation and trains a separate supervised linear probe. All V3 samplers are uniform; auxiliary class weights use training labels only.

Model	U	T	I	Input	Pairing	Representation	Dimension
Dual_U0T0I0	0	0	0	dual, 8 kHz	same-coda augmented views	trained, 50 epochs	64
Dual_U0T0I1	0	0	1	dual, 8 kHz	same-coda augmented views	trained, 50 epochs	64
Dual_U0T1I0	0	1	0	dual, 8 kHz	same-coda augmented views	trained, 50 epochs	64
Dual_U0T1I1	0	1	1	dual, 8 kHz	same-coda augmented views	trained, 50 epochs	64
Dual_U1T0I0	1	0	0	dual, 8 kHz	same-unit whole codas	trained, 50 epochs	64
Dual_U1T0I1	1	0	1	dual, 8 kHz	same-unit whole codas	trained, 50 epochs	64
Dual_U1T1I0	1	1	0	dual, 8 kHz	same-unit whole codas	trained, 50 epochs	64
Dual_U1T1I1	1	1	1	dual, 8 kHz	same-unit whole codas	trained, 50 epochs	64
rhythm_UTI0	1	1	0	rhythm, 8 kHz	same-unit whole codas	trained, 50 epochs	64
rhythm_UTI1	1	1	1	rhythm, 8 kHz	same-unit whole codas	trained, 50 epochs	64
spectral_UTI0	1	1	0	spectral, 8 kHz	same-unit whole codas	trained, 50 epochs	64
spectral_UTI1	1	1	1	spectral, 8 kHz	same-unit whole codas	trained, 50 epochs	64
concat_UTI0	1	1	0	concat, 8 kHz	same-unit whole codas	trained, 50 epochs	128
concat_UTI1	1	1	1	concat, 8 kHz	same-unit whole codas	trained, 50 epochs	128
Dual_swap_UTI	1	1	1	dual, 8 kHz	same-unit swapped channels	trained, 50 epochs	64
Dual_16k_UTI0	1	1	0	dual, 16 kHz	same-unit whole codas	trained, 50 epochs	64
Dual_16k_UTI1	1	1	1	dual, 16 kHz	same-unit whole codas	trained, 50 epochs	64
AVES_FT	CE	CE	CE	audio	none	transformer fine-tuned, 10 epochs; frontend frozen	64
AVES_ICI_FT	CE	CE	CE	audio + ICI	none	transformer fine-tuned, 10 epochs; frontend frozen	64

Single-channel auxiliary heads both use the available channel. Dual/concatenation type heads use rhythm and ID heads use spectrum. U=0 removes unit labels from representation training but not from evaluation partitioning/probes. Label-free objectives still use annotation-derived ICI inputs. AVES FT has a different pretraining history, parameter count and epoch budget from DCCE.

V2 benchmark: downstream whale-label access

These columns describe local downstream training, not all labels used in external pretraining. Every row uses a subsequently trained supervised probe.

Model	Unit use	Type loss	ID loss	Unit-balanced sampler	Representation	Input
Raw ICI	none	no	no	no	fixed features	annotated ICI
Temporal mel	none	no	no	no	fixed features	audio
ICI + temporal mel	none	no	no	no	fixed features	annotated ICI + audio
WhAM	none	no	no	no	frozen pretrained; WhAM/AVES layer selected on validation	audio
AVES	none	no	no	no	frozen pretrained; WhAM/AVES layer selected on validation	audio
BirdNET	none	no	no	no	frozen pretrained; WhAM/AVES layer selected on validation	audio
DCCE no ID	whole-coda unit pairing	yes	no	yes	trained, 50 epochs	annotated ICI + audio
DCCE ID	whole-coda unit pairing	yes	yes	yes	trained, 50 epochs	annotated ICI + audio
DCCE swap ID	same-unit swapped channels	yes	yes	yes	trained, 50 epochs	annotated ICI + audio
Supervised dual	unit CE	yes	yes	yes	trained head/fusion; AVES backbone frozen in adapter rows	annotated ICI + audio
AVES adapter	unit CE	yes	yes	yes	trained head/fusion; AVES backbone frozen in adapter rows	audio
AVES + ICI adapter	unit CE	yes	yes	yes	trained head/fusion; AVES backbone frozen in adapter rows	annotated ICI + audio

Additional frozen input-matched controls

WhAM + ICI, AVES + ICI, and BirdNET + ICI concatenate annotated timing with frozen audio embeddings. They use no local representation-training labels, losses, pairing, or sampler. Only the training-only scaler and supervised linear probe are fitted; WhAM/AVES layers are selected on validation data. External pretraining label/data access remains as described above. These controls are evaluated on all six random/date partitions and all four year holdouts.

The strict year diagnostic reuses the same definitions for Raw ICI, ICI + temporal mel, WhAM, AVES, DCCE 110 and DCCE 111. The latter two are trained afresh on each year partition with uniform sampling. Saved-branch probes are post-hoc diagnostics on fixed DCCE 110/111 checkpoints and do not replace the primary fused-representation results.

summary = pd.read_csv(OUT / 'summary.csv')
display(summary.pivot_table(index=['version','protocol','model'], columns='task', values=['mean','std','count']))

			count			mean			std
		task	coda_type	individual_id	unit	coda_type	individual_id	unit	coda_type	individual_id	unit
version	protocol	model
v2	date_grouped	AVES	3.0	3.0	3.0	0.061377	0.073839	0.699606	0.022273	0.051010	0.099810
		AVES_Adapter	3.0	3.0	3.0	0.078889	0.058777	0.646327	0.028765	0.043898	0.134228
		AVES_ICI_Adapter	3.0	3.0	3.0	0.183220	0.081183	0.668603	0.037160	0.072343	0.118781
		BirdNET	3.0	3.0	3.0	0.062522	0.084921	0.611727	0.010718	0.021797	0.133217
		DCCE_ID	3.0	3.0	3.0	0.502319	0.257250	0.573893	0.017592	0.062791	0.074437
		DCCE_no_ID	3.0	3.0	3.0	0.615727	0.281170	0.686573	0.065519	0.108523	0.023382
		DCCE_swap_ID	3.0	3.0	3.0	0.475978	0.219252	0.611110	0.084046	0.088875	0.014209
		ICI_Temporal_Mel	3.0	3.0	3.0	0.667575	0.255334	0.607584	0.050841	0.061528	0.149813
		Raw_ICI	3.0	3.0	3.0	0.956408	0.379035	0.539629	0.036222	0.102940	0.035826
		Supervised_Dual	3.0	3.0	3.0	0.675403	0.269741	0.611473	0.057831	0.095472	0.079680
		Temporal_Mel	3.0	3.0	3.0	0.068907	0.129634	0.568908	0.011673	0.074012	0.094056
		WhAM	3.0	3.0	3.0	0.065956	0.100725	0.666666	0.016308	0.074519	0.155343
	random	AVES	3.0	3.0	3.0	0.252601	0.491131	0.909781	0.027528	0.029407	0.009811
		AVES_Adapter	3.0	3.0	3.0	0.271576	0.478530	0.932544	0.027186	0.018312	0.009488
		AVES_ICI_Adapter	3.0	3.0	3.0	0.451531	0.582073	0.939995	0.023057	0.078058	0.009122
		BirdNET	3.0	3.0	3.0	0.215446	0.480356	0.887335	0.052697	0.021991	0.019882
		DCCE_ID	3.0	3.0	3.0	0.721958	0.588005	0.889366	0.051336	0.079361	0.004248
		DCCE_no_ID	3.0	3.0	3.0	0.799952	0.626933	0.906846	0.026307	0.056997	0.002644
		DCCE_swap_ID	3.0	3.0	3.0	0.636423	0.640799	0.879231	0.042360	0.065771	0.028993
		ICI_Temporal_Mel	3.0	3.0	3.0	0.804218	0.723783	0.837426	0.015689	0.039093	0.012893
		Raw_ICI	3.0	3.0	3.0	0.987453	0.439054	0.570300	0.008672	0.007079	0.023350
		Supervised_Dual	3.0	3.0	3.0	0.761745	0.683831	0.885832	0.119211	0.080852	0.016999
		Temporal_Mel	3.0	3.0	3.0	0.201753	0.429238	0.772509	0.024982	0.044984	0.006913
		WhAM	3.0	3.0	3.0	0.258139	0.449791	0.905812	0.045638	0.051887	0.011270
v3	date_grouped	AVES_FT	3.0	3.0	3.0	0.071606	0.059693	0.626429	0.050473	0.026242	0.157076
		AVES_ICI_FT	3.0	3.0	3.0	0.210730	0.072674	0.640972	0.068920	0.038355	0.157738
		Dual_16k_UTI0	3.0	3.0	3.0	0.708279	0.249994	0.624067	0.146885	0.131308	0.003592
		Dual_16k_UTI1	3.0	3.0	3.0	0.669489	0.172909	0.611894	0.011272	0.141528	0.055992
		Dual_U0T0I0	3.0	3.0	3.0	0.468995	0.181433	0.541650	0.042615	0.133170	0.090488
		Dual_U0T0I1	3.0	3.0	3.0	0.441014	0.158994	0.484759	0.055972	0.054353	0.101975
		Dual_U0T1I0	3.0	3.0	3.0	0.662280	0.208878	0.551268	0.080695	0.090873	0.078475
		Dual_U0T1I1	3.0	3.0	3.0	0.679337	0.173883	0.551426	0.036979	0.068378	0.113046
		Dual_U1T0I0	3.0	3.0	3.0	0.348014	0.244719	0.517787	0.039266	0.084209	0.139986
		Dual_U1T0I1	3.0	3.0	3.0	0.270115	0.133899	0.511206	0.010946	0.038532	0.117560
		Dual_U1T1I0	3.0	3.0	3.0	0.692072	0.248122	0.597502	0.088280	0.129935	0.077931
		Dual_U1T1I1	3.0	3.0	3.0	0.627668	0.223143	0.595997	0.074429	0.175119	0.052135
		Dual_swap_UTI	3.0	3.0	3.0	0.471134	0.189278	0.621251	0.098548	0.069411	0.070424
		concat_UTI0	3.0	3.0	3.0	0.859337	0.275378	0.549133	0.014456	0.116570	0.171126
		concat_UTI1	3.0	3.0	3.0	0.849676	0.239112	0.527221	0.082756	0.042940	0.140180
		rhythm_UTI0	3.0	3.0	3.0	0.927080	0.352297	0.486565	0.046769	0.133690	0.090828
		rhythm_UTI1	3.0	3.0	3.0	0.918903	0.330671	0.505141	0.044746	0.143114	0.113215
		spectral_UTI0	3.0	3.0	3.0	0.062041	0.053608	0.540588	0.025225	0.046902	0.129682
		spectral_UTI1	3.0	3.0	3.0	0.062850	0.091767	0.558569	0.011001	0.016134	0.093290
	random	AVES_FT	3.0	3.0	3.0	0.243197	0.414024	0.923969	0.025432	0.078094	0.006931
		AVES_ICI_FT	3.0	3.0	3.0	0.466130	0.543899	0.912700	0.058863	0.016969	0.010566
		Dual_16k_UTI0	3.0	3.0	3.0	0.911201	0.630982	0.901471	0.057115	0.046188	0.006359
		Dual_16k_UTI1	3.0	3.0	3.0	0.727462	0.623031	0.893116	0.088185	0.045712	0.007781
		Dual_U0T0I0	3.0	3.0	3.0	0.582662	0.623958	0.693250	0.041934	0.064456	0.026777
		Dual_U0T0I1	3.0	3.0	3.0	0.565726	0.592438	0.673675	0.016813	0.061273	0.015982
		Dual_U0T1I0	3.0	3.0	3.0	0.769836	0.591471	0.709490	0.085769	0.057232	0.026989
		Dual_U0T1I1	3.0	3.0	3.0	0.801521	0.600153	0.725360	0.045655	0.080143	0.013393
		Dual_U1T0I0	3.0	3.0	3.0	0.492673	0.434126	0.861146	0.090645	0.066791	0.001385
		Dual_U1T0I1	3.0	3.0	3.0	0.348301	0.474068	0.885048	0.111685	0.031556	0.018935
		Dual_U1T1I0	3.0	3.0	3.0	0.865400	0.597034	0.900959	0.114735	0.051078	0.020301
		Dual_U1T1I1	3.0	3.0	3.0	0.735336	0.580833	0.907371	0.090526	0.011902	0.010447
		Dual_swap_UTI	3.0	3.0	3.0	0.708329	0.657122	0.853837	0.073161	0.066612	0.012703
		concat_UTI0	3.0	3.0	3.0	0.938746	0.635740	0.875769	0.062073	0.015360	0.024578
		concat_UTI1	3.0	3.0	3.0	0.926869	0.695642	0.872877	0.059922	0.054608	0.010124
		rhythm_UTI0	3.0	3.0	3.0	0.978318	0.477465	0.581279	0.020882	0.036741	0.043923
		rhythm_UTI1	3.0	3.0	3.0	0.973165	0.454129	0.604256	0.012880	0.060516	0.052289
		spectral_UTI0	3.0	3.0	3.0	0.202073	0.357716	0.875258	0.021276	0.103299	0.022978
		spectral_UTI1	3.0	3.0	3.0	0.199069	0.444400	0.853682	0.037031	0.092184	0.006178

Conditional uncertainty

Date-cluster bootstrap confidence intervals condition on each fitted model. They do not capture training variability. Intervals are not estimated for fewer than two supported test dates. Classes absent from a bootstrap resample score zero under the fixed original class set. Three-run SD and per-run intervals answer different questions. Full intervals are in results_v3/cluster_bootstrap.csv; secondary scores including unseen classes are in results_v3/secondary_summary_including_unseen.csv.

intervals = pd.read_csv(OUT / 'cluster_bootstrap.csv')
display(intervals)

	model	protocol	seed	version	task	lower	upper	n_dates	n_test
0	AVES	date_grouped	0	v2	unit	0.282067	0.620450	5	328
1	AVES	date_grouped	0	v2	coda_type	0.007123	0.047641	5	326
2	AVES	date_grouped	0	v2	individual_id	NaN	NaN	1	186
3	AVES_Adapter	date_grouped	0	v2	unit	0.235757	0.548072	5	328
4	AVES_Adapter	date_grouped	0	v2	coda_type	0.011496	0.057891	5	326
...	...	...	...	...	...	...	...	...	...
553	spectral_UTI0	random	2	v3	coda_type	0.101263	0.224341	40	277
554	spectral_UTI0	random	2	v3	individual_id	0.138896	0.340289	23	151
555	spectral_UTI1	random	2	v3	unit	0.746105	0.904120	40	277
556	spectral_UTI1	random	2	v3	coda_type	0.097461	0.197164	40	277
557	spectral_UTI1	random	2	v3	individual_id	0.163604	0.449678	23	151

558 rows × 9 columns

within = pd.read_csv(OUT / 'within_unit_id.csv')
display(within)
print('Blank scores indicate fewer than two supported train/test identities.')

	model	protocol	seed	version	unit	n_train	n_test	n_train_ids	n_test_ids	macro_f1	selected_layer	n_validation	validation_macro_f1	probe_convergence_warning
0	AVES	date_grouped	0	v2	A	0	0	0	0	NaN	NaN	NaN	NaN	NaN
1	AVES	date_grouped	0	v2	D	10	0	2	0	NaN	NaN	NaN	NaN	NaN
2	AVES	date_grouped	0	v2	F	420	186	7	3	0.019608	10.0	111.0	0.257803	False
3	AVES_Adapter	date_grouped	0	v2	A	0	0	0	0	NaN	NaN	NaN	NaN	NaN
4	AVES_Adapter	date_grouped	0	v2	D	10	0	2	0	NaN	NaN	NaN	NaN	NaN
...	...	...	...	...	...	...	...	...	...	...	...	...	...	...
553	spectral_UTI0	random	2	v3	D	6	2	2	2	0.333333	0.0	2.0	1.000000	False
554	spectral_UTI0	random	2	v3	F	458	144	7	7	0.343738	0.0	115.0	0.379803	False
555	spectral_UTI1	random	2	v3	A	18	5	2	2	0.583333	0.0	4.0	0.733333	False
556	spectral_UTI1	random	2	v3	D	6	2	2	2	0.333333	0.0	2.0	0.000000	False
557	spectral_UTI1	random	2	v3	F	458	144	7	7	0.370712	0.0	115.0	0.411015	False

558 rows × 14 columns

Blank scores indicate fewer than two supported train/test identities.

for protocol in ['random','date_grouped']:
    figure = OUT / 'figures' / f'{protocol}.png'
    if figure.exists():
        display(Image(filename=str(figure)))

verification = OUT / 'verification.json'
if verification.exists():
    display(json.loads(verification.read_text()))
else:
    print('Full verification is pending completion of the experiment grid.')
preserved = ROOT / 'registry_v3/v2_preservation.json'
display(json.loads(preserved.read_text()))

{'checked': 114,
 'expected': 114,
 'missing': [],
 'all_complete': True,
 'checks': 'source and split hashes; train/test separation; train-only scaler; uniform sampler metadata; fixed epochs; finite losses and embeddings; supported labels and recomputed scores; probe convergence'}

{'checked': 516, 'changed': []}

year_report = ROOT / 'reports_v3/year_holdout.md'
if year_report.exists():
    display(Markdown(year_report.read_text()))
    display(json.loads((OUT / 'year_holdout_verification.json').read_text()))
else:
    print('Year-holdout diagnostic is queued after the main grid.')

Explicit year holdout diagnostic

This exploratory amendment was specified after partial V3 random-split results were visible, before any year-holdout execution. Each year is held out completely; training and validation use disjoint dates from other years. All six prespecified models and all four years are reported. Neural seed is fixed at 0, so these are four different populations, not four repeat-seed estimates.

Partition sizes

Held-out year	Train codas	Validation codas	Test codas
2005	687	328	368
2008	845	275	263
2009	967	321	95
2010	570	156	657

unit

Model	2005	2008	2009	2010
Raw ICI	0.224 (368/368 supported)	0.425 (263/263 supported)	0.541 (95/95 supported)	0.476 (657/657 supported)
ICI + temporal mel	0.649 (368/368 supported)	0.360 (263/263 supported)	0.778 (95/95 supported)	0.327 (657/657 supported)
WhAM	0.822 (368/368 supported)	0.371 (263/263 supported)	0.789 (95/95 supported)	0.422 (657/657 supported)
AVES	0.750 (368/368 supported)	0.510 (263/263 supported)	0.833 (95/95 supported)	0.377 (657/657 supported)
DCCE 110	0.592 (368/368 supported)	0.362 (263/263 supported)	0.767 (95/95 supported)	0.346 (657/657 supported)
DCCE 111	0.448 (368/368 supported)	0.391 (263/263 supported)	0.713 (95/95 supported)	0.357 (657/657 supported)

coda_type

Model	2005	2008	2009	2010
Raw ICI	0.957 (358/368 supported)	0.908 (263/263 supported)	1.000 (95/95 supported)	0.965 (636/657 supported)
ICI + temporal mel	0.589 (358/368 supported)	0.607 (263/263 supported)	0.977 (95/95 supported)	0.427 (636/657 supported)
WhAM	0.077 (358/368 supported)	0.045 (263/263 supported)	0.166 (95/95 supported)	0.051 (636/657 supported)
AVES	0.065 (358/368 supported)	0.047 (263/263 supported)	0.103 (95/95 supported)	0.029 (636/657 supported)
DCCE 110	0.596 (358/368 supported)	0.702 (263/263 supported)	1.000 (95/95 supported)	0.476 (636/657 supported)
DCCE 111	0.583 (358/368 supported)	0.546 (263/263 supported)	0.839 (95/95 supported)	0.421 (636/657 supported)

individual_id

Model	2005	2008	2009	2010
Raw ICI	0.404 (236/314 supported)	0.542 (57/57 supported)	0.412 (37/64 supported)	0.232 (292/319 supported)
ICI + temporal mel	0.353 (236/314 supported)	0.491 (57/57 supported)	0.410 (37/64 supported)	0.128 (292/319 supported)
WhAM	0.238 (236/314 supported)	0.229 (57/57 supported)	0.038 (37/64 supported)	0.093 (292/319 supported)
AVES	0.193 (236/314 supported)	0.207 (57/57 supported)	0.228 (37/64 supported)	0.068 (292/319 supported)
DCCE 110	0.258 (236/314 supported)	0.607 (57/57 supported)	0.468 (37/64 supported)	0.197 (292/319 supported)
DCCE 111	0.257 (236/314 supported)	0.670 (57/57 supported)	0.349 (37/64 supported)	0.186 (292/319 supported)

Full-eligible macro-F1 values (including unseen labels) are in year_holdout_metrics.csv. Conditional date-cluster intervals are in year_holdout_intervals.csv. Class support and composition differ across years. These tests remove shared years between train/validation and test, but performance differences still combine biological, recording and label-composition changes; they do not identify a causal recording artifact.

{'checked': 24, 'expected': 24, 'all_complete': True}

for report in ['feature_fusion','branch_probes']:
    path = ROOT / 'reports_v3' / f'{report}.md'
    if path.exists():
        display(Markdown(path.read_text()))

Input-matched frozen-feature controls

Exploratory addition: all three frozen pretrained families receive the same annotated ICI through direct concatenation. Only the downstream linear probe is trained. Multi-layer selection uses validation data exclusively. These controls match available inputs; they do not match external pretraining history or local auxiliary-label objectives.

date_grouped

Model	Unit	Type	ID
AVES_ICI	0.712 ± 0.111	0.366 ± 0.084	0.128 ± 0.103
BirdNET_ICI	0.628 ± 0.095	0.287 ± 0.061	0.100 ± 0.055
WhAM_ICI	0.768 ± 0.057	0.376 ± 0.151	0.153 ± 0.135

random

Model	Unit	Type	ID
AVES_ICI	0.926 ± 0.011	0.667 ± 0.044	0.652 ± 0.044
BirdNET_ICI	0.884 ± 0.018	0.497 ± 0.028	0.580 ± 0.031
WhAM_ICI	0.921 ± 0.005	0.722 ± 0.027	0.598 ± 0.028

Year holdouts

Model	Year	Unit	Type	ID
AVES_ICI	2005	0.675	0.440	0.263
AVES_ICI	2008	0.498	0.332	0.330
AVES_ICI	2009	0.835	0.733	0.376
AVES_ICI	2010	0.354	0.303	0.057
BirdNET_ICI	2005	0.646	0.238	0.229
BirdNET_ICI	2008	0.355	0.239	0.195
BirdNET_ICI	2009	0.819	0.718	0.126
BirdNET_ICI	2010	0.323	0.330	0.101
WhAM_ICI	2005	0.773	0.406	0.305
WhAM_ICI	2008	0.410	0.367	0.459
WhAM_ICI	2009	0.856	0.961	0.255
WhAM_ICI	2010	0.451	0.382	0.146

All per-run support counts and secondary scores are saved in feature_fusion_metrics.csv. Year holdouts control local probe training, not unknown historical pretraining exposure.

Descriptive channel probes

Post-hoc analysis of 12 fixed checkpoints. Encoders are not retrained. Every channel uses the same supervised, training-only frozen-probe protocol as fusion. Re-extracted fusion embeddings are checked against saved embeddings. These are diagnostic scores, not replacements selected for the primary result.

date_grouped: Dual_U1T1I0

Output	Unit	Type	ID
rhythm	0.558 ± 0.091	0.917 ± 0.065	0.325 ± 0.144
spectral	0.516 ± 0.149	0.057 ± 0.023	0.098 ± 0.069
fusion	0.598 ± 0.078	0.692 ± 0.088	0.248 ± 0.130

date_grouped: Dual_U1T1I1

Output	Unit	Type	ID
rhythm	0.552 ± 0.101	0.911 ± 0.071	0.320 ± 0.150
spectral	0.511 ± 0.125	0.055 ± 0.032	0.074 ± 0.045
fusion	0.596 ± 0.052	0.628 ± 0.074	0.223 ± 0.175

random: Dual_U1T1I0

Output	Unit	Type	ID
rhythm	0.639 ± 0.027	0.983 ± 0.023	0.459 ± 0.052
spectral	0.863 ± 0.013	0.158 ± 0.023	0.367 ± 0.116
fusion	0.901 ± 0.020	0.865 ± 0.115	0.597 ± 0.051

random: Dual_U1T1I1

Output	Unit	Type	ID
rhythm	0.632 ± 0.028	0.982 ± 0.019	0.461 ± 0.054
spectral	0.861 ± 0.001	0.169 ± 0.021	0.443 ± 0.059
fusion	0.907 ± 0.010	0.735 ± 0.091	0.581 ± 0.012

Reproduction

Run scripts_v3/experiments.py for the 102 DCCE/control evaluations and scripts_v3/aves_finetune.py for the 12 pretrained-transformer adaptations. scripts_v3/year_holdout.py adds the 24 strict year-holdout evaluations. Runners verify existing source hashes before skipping completed outputs. Analysis scripts regenerate tables and confidence intervals. Consult README_v3.md; training is intentionally not an automatic notebook side effect.