Interpretability Audits of Computer Vision Anemia Screeners Under Conjunctival Pallor Photography Protocol Drift: A Healthcare Calibration Fairness Evaluation

Conjunctival pallor photography is increasingly proposed as a low-cost triage signal for anemia screening, yet the interpretability of computer vision models trained on such images is rarely audited against the photographic protocols under which they are deployed. The framework formalises protocol drift as a shift in the joint distribution of illumination, white balance, eyelid retraction and conjunctival exposure, and derives a decomposition of calibration error into a protocol-invariant component and a drift-sensitive component. A before-after-control-impact design is characterised analytically under a declared variance structure with intraclass correlation 0.64, between-cluster SD 6.0 and residual SD 4.5; 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 design requires 54 clusters to detect a 5-percentage-point effect at 80% power. The framework proposes that interpretability audits report drift-conditional calibration, subgroup fairness gaps and attribution stability as separate quantities, and it specifies the estimands and exhibits an audit would require. The contribution is a methodological specification, not an empirical finding: it identifies which interpretability claims survive protocol drift and which do not, and it argues that fairness evaluation of anemia screeners is uninterpretable without explicit conditioning on acquisition protocol.