Anatomy-Quality-Patient Reliability auditing for fundus-based glaucoma screening under external domain shift

High internal accuracy in fundus-based glaucoma screening does not establish anatomical plausibility, calibration, patient consistency, or transportability. We present Anatomy-Quality-Patient Reliability (AQPR), a structured, retraining-free post-hoc audit framework that integrates established reliability tests for frozen glaucoma classifiers. Models trained on AIROGS-light were externally audited on HYGD and source-wise on SMDG using optic disc/optic cup proxy segmentation, matched-random saliency controls, insertion/deletion, perturbation, quality stratification, patient aggregation, calibration, and shortcut forensics. AQPR revealed architecture-dependent reliability: ConvNeXt-Tiny showed stronger bounded convergence, whereas EfficientNet-B0 showed mixed functional evidence. Source-level calibration remained heterogeneous, reframing trustworthy glaucoma AI as a multi-axis deployment audit.