Evaluation of Concept-Probe Graph-Connectivity Audits for Computer Vision Agriculture Disease Classifiers Under Intercropped Plot Layouts: An Interpretability Segmentation Bound for Smallholder Fungicide Zoning

Concept-probe audits are increasingly used to certify that computer vision disease classifiers attend to lesions rather than to intercropping context, yet the audit's own statistical behaviour under realistic plot layouts has not been characterised. the present analysis treats the concept-probe graph-connectivity audit as an estimator and derives its operating characteristics under a declared variance structure. The audit is formalised as a graph-connectivity functional on the probe-activation matrix, and a segmentation bound is derived that links the audit's false-certification rate to the intraclass correlation of probe activations across intercropped sub-plots. Under a before-after-control-impact design with between-cluster SD = 6.0, residual SD = 4.5 and intraclass correlation = 0.64, power was estimated across 2000 design cells at cluster counts of 8, 16, 24 and 32 and effect sizes of 3, 5 and 8 percentage points. The design required 54 clusters to detect a 5 percentage point effect at 80% power. The bound implies that audits calibrated on monocrop imagery will over-certify classifiers deployed on intercropped smallholder plots whenever the sub-plot intraclass correlation exceeds the audit's assumed value, and that the resulting segmentation error propagates into fungicide zoning as systematic over-application on the intercropped fraction. The contribution is a formal statement of the audit's validity conditions and a design rule for sizing concept-probe audits in intercropped smallholder settings.