UNTESTED registered EVALUATION axis A12 (control condition: `zero-shot audio-LLM`), paired across A3 values no run
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.
The gap between a supervised probe and a zero-shot audio-LLM is a cheap, model-free estimate of how much leakage a protocol admits, because a zero-shot model has no training split and therefore cannot leak speaker identity, demographics, or preprocessing statistics.
Shahin, Ahmed, Epps, 2025 (arXiv:2506.17351) report Qwen2-Audio zero-shot cognitive-impairment detection "comparable to supervised methods" and frame it as a convenience result, not as a measurement instrument. Cross-check required against Kabir & Munira, 2026 (arXiv:2605.24806): handcrafted features + a text LLM are more reliable than raw audio for low-resource languages, so the zero-shot reference must be reported in both input formats or it is language-confounded.
Compute Δ_leaky = AUC_sup(random_recording split) − AUC_zeroshot and Δ_honest = AUC_sup(speaker_disjoint + demographically_matched) − AUC_zeroshot. If Δ_leaky ≈ Δ_honest (overlapping bootstrap CIs) on ≥ 2 of 3 datasets, the proposed leakage estimator is falsified — the gap tracks supervised capacity, not leakage. If Δ_leaky ≫ Δ_honest consistently across ≥ 3 datasets, the field gains a leakage audit that needs no re-splitting of the original data.
Δ_leaky − Δ_honest ∈ [0.08, 0.25] AUC, positive, on SVD and Coswara. Zero-shot absolute AUC is expected to be low (0.55–0.68) — the hypothesis is about the gap, not the zero-shot number.
| tier (as registered) | SCREENING at registration (n ≤ 3), with a pre-declared promotion condition: promote to EVALUATION at n = 10, m = 3 only if zero-shot throughput permits ≥ 10 full passes per dataset inside the foreground GPU windows available. If it does not, V5 ships as screening and is labelled as such — it is never re-described as a result. |
| family size m / seeds n | screening n ≤ 3 · promotion target m = 3 (3 datasets), n = 10 → 0.001953 ≤ 0.016667 ✓ (n = 8 would also satisfy m = 3) |
| datasets | SVD · Coswara · COUGHVID |
| cost | The only real GPU cost in the registry. 4-bit Qwen2-Audio-7B inference; one foreground eval window per dataset, N capped per condition. Host RAM, not VRAM, is the binding constraint. |
| min attainable paired p | 0.001953 |
| Holm-tightest threshold 0.05/m | 0.016667 |
| arithmetically satisfiable under R6 | ✓ |
| registry status | UNTESTED |
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.
UNTESTED No artifact exists for this hypothesis. Under R7 a hypothesis whose falsifier has not been executed is UNTESTED and is never reported as supported. This page exists so that the prediction above is on the record before any result is, and so the debt is visible rather than absent. It is counted as debt on the master dashboard and in the registry table.