WavLM-base+ on full-corpus SVD vs its own confound battery

EVALUATION NOT CLEARED SCOPE-LIMITED svd wavlm

Scope caveat. This run is a pilot slice -- 1,679 of the 13,535 speakers measured elsewhere on svd. It may satisfy the repeat contract and still be a narrower measurement than a later run on the same corpus; both facts are shown because either alone would mislead.

0.8650

speaker-level ROC-AUC, ens_rank3

95% CI [0.847, 0.881]

0.8649

confound bar (age_sex_duration_rms)

n = 8

repeats × 5 folds

n=8 repeats

1,679

speakers

28,509 recordings

Power contract (R6)

Paired Wilcoxon over 8 repeats gives a minimum attainable p of 0.007812 = 2/2⁸, against a Holm-tightest threshold of 0.025000 = 0.05/2 for the pre-registered family. Feasible. This is arithmetic on the design, computed before the data: it says what the run could have shown, not what it did.

Every head and every confound baseline

head / baseline	speaker AUC	95% CI	recording AUC	UAR	ECE	per-repeat mean ± std	verdict
`logreg`	0.8671	[0.850, 0.883] n_boot=2000	0.7438	0.6830	0.145	— ± —	NOT CLEARED
`ens_rank3`	0.8650	[0.847, 0.881] n_boot=2000	0.7443	0.7676	0.157	— ± —	NOT CLEARED
confound `age_only`	0.8642	[0.846, 0.882] n_boot=2000	0.8737	0.7976	0.075	— ± —	baseline
confound `sex_only`	0.5027	[0.474, 0.532] n_boot=2000	0.4898	0.5000	0.067	— ± —	baseline
confound `age_sex`	0.8646	[0.847, 0.882] n_boot=2000	0.8743	0.8045	0.087	— ± —	baseline
confound `duration_only`	0.5021	[0.474, 0.531] n_boot=2000	0.4724	0.5000	0.068	— ± —	baseline
confound `intensity_rms_only`	0.5206	[0.492, 0.549] n_boot=2000	0.5121	0.5000	0.068	— ± —	baseline
confound `age_sex_duration_rms`	0.8649	[0.847, 0.882] n_boot=2000	0.8747	0.8032	0.090	— ± —	THE BAR

ECE is expected calibration error: the gap between stated confidence and observed frequency. It is reported here because clinical usability is a probability question, not a ranking question (CLAUDE.md §4.3). A head can rank well and still be badly calibrated — compare the ensemble's ECE against the plain logistic head.

Margin against the confound bar, per head

head	recording-level delta [95% CI]	speaker-level delta [95% CI]	paired Wilcoxon over repeats	cleared?
`logreg`	-0.1310 [-0.147, -0.115], p>0 = 0.0000	+0.0022 [-0.015, 0.020], p>0 = 0.5845	W = 0.0, p = 0.007812	NO
`ens_rank3`	-0.1305 [-0.147, -0.114], p>0 = 0.0000	+0.0001 [-0.017, 0.018], p>0 = 0.5030	W = 0.0, p = 0.007812	NO

Per-repeat traces

head / baseline	AUC per repeated partition
`logreg`	0.7386 · 0.7382 · 0.7367 · 0.7413 · 0.7380 · 0.7355 · 0.7395 · 0.7367
`ens_rank3`	0.7386 · 0.7379 · 0.7374 · 0.7411 · 0.7382 · 0.7359 · 0.7397 · 0.7369
`confound::age_only`	0.8744 · 0.8741 · 0.8738 · 0.8744 · 0.8745 · 0.8740 · 0.8732 · 0.8745
`confound::sex_only`	0.5293 · 0.5261 · 0.5097 · 0.5277 · 0.5285 · 0.5249 · 0.5285 · 0.5218
`confound::age_sex`	0.8751 · 0.8743 · 0.8732 · 0.8754 · 0.8751 · 0.8746 · 0.8735 · 0.8751
`confound::duration_only`	0.4859 · 0.4920 · 0.4908 · 0.4936 · 0.4919 · 0.4836 · 0.4924 · 0.4901
`confound::intensity_rms_only`	0.5090 · 0.5114 · 0.5129 · 0.5138 · 0.5119 · 0.5110 · 0.5104 · 0.5125
`confound::age_sex_duration_rms`	0.8752 · 0.8744 · 0.8733 · 0.8753 · 0.8751 · 0.8748 · 0.8736 · 0.8751

The spread across repeated speaker-disjoint partitions is the empirical noise band. A delta smaller than this spread is not a result whatever its point estimate says.

Reproduction provenance

artifact	`autoresearch_results/bench_svd_wavlm_mean_std.json` · view raw · last written 2026-07-28 04:47 UTC
command	`run_benchmark.py --corpus svd --backbone wavlm --head logreg --folds 5 --repeats 8`
git_sha	`5cab307b744df840d8a16ac305c47192e2ab0aaf`
config_hash	`815673703168e601`
generated_utc	`2026-07-28T04:47:05Z`
embedding_content_hash	`b2bea8e432783ff4c137d640aea10459092311b73a7a30a7b41a252806c80e6c`
wall clock	1,282.4 s (0.36 h)
written up as	F2, F3 in FINDINGS.md

What this page does not have

No 7-step reasoning entry exists for this run. The sibling convention renders Diagnosis · Citations · Hypothesis · Prediction · Verdict · Learning on every per-experiment page, sourced from a reasoning_annotations.json. This program never wrote one, and neither does it keep an append-only experiment_log.jsonl; the run history is the set of artifact files listed on the run ledger, ordered only by their write time. The pre-registered prediction for this run does exist — it is on the hypothesis registry, but this run predates the registry and has no pre-registered prediction of its own. Everything else on that convention’s checklist is absent here and is shown as absent rather than quietly dropped.