Interpretability Audits of Computer Vision Prescription Pill Identification Models Under Pharmacy Refill Label Reprint Variability: A Healthcare Fairness Estimation

Computer vision models that identify prescription pills from smartphone photographs are increasingly proposed as adherence aids, yet their interpretability audits typically assume that the label text printed on a pharmacy refill container is stable across reprints. the present analysis treats reprint variability as a structured nuisance that can distort both model attention and the fairness estimates derived from that attention. Because no observational corpus of reprint-degraded labels is available, the study is framed as a methodological and theoretical analysis rather than an empirical audit. It develops a formal account of reprint-induced distribution shift, defines an interpretability-audit estimand that separates label-content attention from reprint-artefact attention, and characterises the operating characteristics of a before-after-control-impact design for estimating reprint effects on subgroup fairness. Under a declared variance structure with between-cluster SD 6.0, residual SD 4.5 and intraclass correlation 0.64, power was estimated across 2000 design cells for cluster counts of 8, 16, 24 and 32 and effect sizes of 3, 5 and 8 percentage points. The design requires 54 clusters to detect a 5-percentage-point effect at 80% power. Interpretability audits of pill-identification models are only informative when reprint variability is treated as a design parameter rather than as a post hoc robustness check, and it specifies the audit protocol that follows from that position.