Evaluation of Concept-Probe Compositional-Generalization Audits for Computer Vision Agriculture Disease Classifiers Under Novel Pathogen-Symptom Combinations: An Interpretability Bound for Smallholder Fungicide Selection

Computer vision classifiers for crop disease are increasingly deployed to guide fungicide selection by smallholder farmers, yet their reliability under novel pathogen-symptom combinations remains uncharacterised. The framework treats a classifier as a composition of concept detectors and a decision head, and defines compositional-generalization gap as the divergence between concept activation patterns on seen versus novel symptom combinations. Using a variance-components model with between-cluster standard deviation 6.0, residual standard deviation 4.5 and intraclass correlation 0.64, the design analysis characterised the operating characteristics of a before-after-control-impact estimator across 2000 replications per design cell for cluster counts of 8, 16, 24 and 32 and effect sizes of 3, 5 and 8 percentage points. Power to detect a 5 percentage-point improvement reached 80% at 54 clusters. The derived bound specifies the maximum tolerable probe infidelity for a given fungicide-selection error rate, and the analysis shows that compositional-generalization audits can certify classifier safety only when probe fidelity exceeds a threshold that depends jointly on the intraclass correlation and the number of audit clusters. The contribution is a formal interpretability bound that converts concept-probe audit outcomes into a defensible constraint on deployment scope for smallholder fungicide advisory systems.