Evaluation of Concept-Probe Spatial-Resolution Sensitivity Audits for Computer Vision Agriculture Disease Classifiers Under Multispectral Band Coarsening: An Interpretability Fidelity Bound for Smallholder Fungicide Triage

Concept-probe audits are increasingly used to certify that computer vision disease classifiers for smallholder crops attend to the lesions they claim to detect, yet the audit's own sensitivity to the spatial resolution at which concept annotations are drawn has not been bounded. the present analysis treats the concept-probe spatial-resolution sensitivity audit as a formal estimator and derives a fidelity bound that relates probe agreement to the coarsening of multispectral bands and the spatial scale of the concept mask. The classifier is modelled as a Lipschitz-continuous scoring function on a band-coarsened input, and the probe as a linear read-out of an intermediate feature map; the bound separates an irreducible concept-scale term from a coarsening term that grows with the ratio of mask cell size to lesion scale. The derivation yields a certified radius within which probe agreement cannot fall below a stated fidelity threshold, and it identifies the band-coarsening ratio at which the audit loses the power to distinguish a lesion-attending classifier from a background-attending one. The bound is stated as a design condition for smallholder fungicide triage, where a false certificate of attention translates into a mis-timed spray recommendation. No observational data are used or claimed; the contribution is a methodological and formal one, and the bound is offered as a specification that future audits can be checked against rather than as an empirical finding.