PARTIAL registered EVALUATION axis A5 (representation → confound-only feature sets), which is also the `AUC_conf_max` battery of `COMPOSITE.md` §3 1 run(s) on disk
Everything in this block was recorded in version control before any run. It is rendered above the result on purpose: a prediction a reader meets after the number it predicted is not a prediction.
Non-phonatory signal alone (silence/pause structure, recording-level duration and intensity statistics) achieves a substantial fraction of headline performance on voice-health corpora outside the one where the effect was first documented.
Liu, Feng, Yuan, Ling, Interspeech 2024, Clever Hans Effect Found in Automatic Detection of Alzheimer's Disease through Speech (arXiv:2406.07410): near-100% AD detection accuracy from silent segments alone in the Pitt corpus, dropping to ~80% on other datasets or preprocessed Pitt. The audited proposition is the implicit assumption that this is a Pitt-specific artifact. audits/NOVELTY_CRITIQUE.md §d.3 warns that re-deriving known Pitt/DAIC-WOZ results is redundant — so this hypothesis deliberately targets corpora that have not been checked, and Pitt itself is DUA-gated for us anyway.
If silence_only and duration+intensity_only baselines both score < 0.60 AUC on all of SVD, Coswara and COUGHVID under A3 = speaker_disjoint, the generalisation claim is falsified and the shortcut is confirmed Pitt-specific for this corpus set.
silence_only AUC ∈ [0.55, 0.70] on Coswara (crowd-recorded, heterogeneous protocol) and ∈ [0.50, 0.60] on SVD (controlled studio protocol, sustained vowels) — i.e. the shortcut is predicted to scale with acquisition heterogeneity, not with disease.
| tier (as registered) | EVALUATION |
| family size m / seeds n | m = 6 (3 corpora × 2 shortcut feature sets) · n = 10 → 0.001953 ≤ 0.008333 ✓ |
| datasets | SVD · Coswara · COUGHVID |
| cost | CPU-only; VAD + functionals. |
| min attainable paired p | 0.001953 |
| Holm-tightest threshold 0.05/m | 0.008333 |
| arithmetically satisfiable under R6 | ✓ |
| registry status | PARTIAL — shortcut does NOT generalise; my mechanism was WRONG — see F6 |
Registry status is read from the ## Summary table of IDEA_TABLE.md, which the author maintains as a set. The per-block status cells are maintained one at a time and disagree with it; one source of truth, and this is it.
EVALUATION n = 10 repeats against the pre-registered family (m = 6 (3 corpora × 2 shortcut feature sets) · n = 10 → 0.001953 ≤ 0.008333 ✓): min attainable paired p = 0.001953 ≤ Holm 0.05/m = 0.008333. Feasible. Feasibility is a property of the design, not of the result — it says what this run could have shown.
EVALUATION PARTIAL -- shortcut does NOT generalise n = 10 × 5 folds
Artifact autoresearch_results/V7_silence_shortcut.json · written 2026-07-28 12:56 UTC · full run page → · written up as F6 in FINDINGS.md
Rendered from autoresearch_results/V7_silence_shortcut.json. Every other run for this hypothesis, including controls, has its own page linked above.
| corpus | features | recordings | speakers | AUC | 95% CI | AUC directionless | status |
|---|---|---|---|---|---|---|---|
| svd | silence_only | 28,509 | 1,679 | 0.5136 | [0.5104, 0.5174] | 0.5136 | RUN |
| svd | duration_intensity_only | 28,509 | 1,679 | 0.5048 | [0.5020, 0.5077] | 0.5048 | RUN |
| coswara | silence_only | 610 | 72 | 0.4854 | [0.4668, 0.5037] | 0.5146 | RUN |
| coswara | duration_intensity_only | not yet measured | not yet measured | not yet measured | not yet measured | not yet measured | NOT RUN -- no data |
| coughvid | silence_only | 13,535 | 13,535 | 0.5264 | [0.5251, 0.5277] | 0.5264 | RUN |
| coughvid | duration_intensity_only | not yet measured | not yet measured | not yet measured | not yet measured | not yet measured | NOT RUN -- no data |
4 of 6 pre-registered cells ran, and the artifact reports its own limitation in a field: falsifier_fully_evaluable = False. all_below_0.60 = True. The corpora measured here are sustained vowels and coughs; Pitt, where the audited shortcut was found, is spontaneous speech in which pause structure carries cognitive load. So this is not evidence that the Pitt effect was spurious — it is evidence that the shortcut does not transfer to these acoustic conditions.