Evaluation of Concept-Probe Spatial-Sensitivity Audits for Computer Vision Agriculture Disease Classifiers Under Sub-Lesion Resolution Limits: An Interpretability Localization Bound for Smallholder Fungicide Placement

Concept-probe audits are increasingly used to certify that computer vision disease classifiers attend to lesion tissue rather than to field context, yet the spatial resolution at which such audits can localise evidence has not been bounded. The bound is derived from a signal-detection model in which probe sensitivity is a function of lesion-to-pixel ratio, probe receptive-field width and classifier confidence margin; it predicts that below a critical lesion-to-pixel ratio the audit cannot distinguish lesion-directed from context-directed attention at any sample size. A before-after-control-impact design was used to characterise the audit's operating characteristics under a declared variance structure (between-cluster SD = 6.0, residual SD = 4.5, intraclass correlation = 0.64), with power 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 achieved 80% power at 56 clusters for a 5-percentage-point effect. The bound implies that audits conducted at resolutions coarser than the lesion scale cannot license pixel-level fungicide placement recommendations, and it specifies the resolution and cluster count at which an audit becomes informative for smallholder spraying decisions.