Evaluation of Counterfactual Concept-Robustness Audits for Computer Vision Agriculture Disease Classifiers Under Camera-Sensor Degradation: An Interpretability Reliability Bound for Smallholder Fungicide Decisions

Computer vision disease classifiers deployed on smallholder smartphones can misclassify when camera sensors degrade, yet interpretability audits rarely quantify how far their explanations can be trusted under such degradation. The audit is formalised as a constrained optimisation over concept activation vectors under a declared sensor-degradation model, and the bound is expressed as a function of the intraclass correlation of concept responses across degradation levels. Using a design-based evaluation with a before-after-control-impact estimator, the operating characteristics of the audit are characterised across cluster counts of 8, 16, 24 and 32 and effect sizes of 3, 5 and 8 percentage points, with 2000 replications per design cell. The intraclass correlation of concept responses is 0.64, and 54 clusters are required to detect a 5 percentage-point effect at 80% power. The bound shows that explanation stability, not classifier accuracy alone, governs the reliability of spray recommendations, and that audits below a threshold cluster count cannot certify decisions for smallholder use.