Interpretability Audits of Computer Vision Wound Infection Classifiers Under Home-Based Smartphone Lighting Variation: A Calibration Fairness Evaluation

Computer vision classifiers that screen smartphone photographs of wounds for infection are increasingly proposed for home use, where illumination is uncontrolled and often poor. the present analysis treats the interpretability audit itself as the object of study and asks what a calibration-fairness audit can and cannot establish when the audit is conducted under the same lighting variation that degrades the classifier. A formal framework is developed in which a classifier's decision function, its post-hoc attribution map, and the audit's calibration estimator are jointly indexed by an illumination nuisance parameter. Three results follow. First, an attribution-stability bound is derived showing that the audit's apparent fairness gap decomposes into a classifier component and an audit component, and that the second does not vanish as sample size grows. Second, a calibration-fairness estimand is defined that is invariant to the audit's own lighting sensitivity, and its identification conditions are stated. Third, the framework is applied to a before-after-control-impact design used to evaluate the audit procedure; under a declared variance structure with intraclass correlation 0.64, between-cluster SD 6.0 and residual SD 4.5, 52 clusters are required for 80% power at a 5 percentage-point effect, and power was estimated across 2000 replications per design cell for cluster counts of 8, 16, 24 and 32 and effects of 3, 5 and 8 percentage points. The contribution is a methodological clarification: interpretability audits of lighting-sensitive classifiers measure the audit as much as the model, and calibration fairness claims require an illumination-invariant estimand rather than a stability heuristic.