V4 — Does calibration degrade *with* distribution shift, or independently of it?

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 claim

Calibration error does not track AUC collapse under corpus shift, so model confidence is unusable as an out-of-domain detector.

The published claim it audits

Kafentzis & Selisios, 2026, Tuberculosis Screening from Cough Audio: Baseline Models, Clinical Variables, and Uncertainty Quantification (arXiv:2601.07969, Sensors 26(4):1223) introduces uncertainty quantification for cough screening and motivates it by the field's protocol heterogeneity. Separately, de Brito, de Souza, Gauy, Finger, Candido Junior, 2025, Fine-tuning Pre-trained Audio Models for COVID-19 Detection (arXiv:2511.14939) documents cross-dataset AUC collapse to 0.43–0.68. Nobody reports cross-corpus ECE / Brier decomposition — the implicit assumption that uncertainty flags shift is untested.

Pre-registered falsifier

Measure ECE, Brier, and reliability curves within- and cross-corpus on COUGHVID↔Coswara (both directions) and on SVD→VOICED. If ECE stays flat within its own seed noise band while AUC falls from ~0.80 to ~0.55, confidence is falsified as a shift detector — a negative result with direct deployment consequences. Raw and temperature-scaled variants reported separately.

Predicted effect (recorded before the run)

AUC falls 0.18–0.30 cross-corpus; raw ECE rises only 0.00–0.04, i.e. inside or barely outside its noise band. Temperature scaling fitted in-domain worsens cross-corpus ECE by 0.02–0.08.