Evaluation of Concept-Probe Sufficiency Audits for Computer Vision Agriculture Disease Classifiers Under Cross-Crop Semantic Shift: An Interpretability Transfer Bound for Smallholder Fungicide Triage

Concept-probe sufficiency audits are increasingly proposed as a precondition for deploying computer-vision disease classifiers in smallholder agriculture, on the assumption that a probe which recovers the diagnostic concept on one crop will recover it on another. the present analysis treats that assumption as a formal object and derives an interpretability transfer bound for it. The audit is modelled as a two-stage estimator: a linear probe fitted on source-crop activations, and a sufficiency test applied to target-crop activations whose concept direction has been rotated by an unobserved semantic shift. Under a declared variance structure (between-cluster SD = 6.0, residual SD = 4.5, intraclass correlation = 0.64), the bound is evaluated 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. The analysis is a design-characterisation study: no observational data are used or claimed, and the reported quantities are the operating characteristics of the audit under the declared variance structure. The principal result is that the sufficiency verdict is not transport-invariant: at the declared intraclass correlation of 0.64, a probe that passes on the source crop can fail on the target crop at effect sizes that the source-crop audit would certify, and 54 clusters are required to detect a 5-percentage-point sufficiency gap at 80% power. The bound reframes concept-probe auditing from a property of a model to a property of a model-crop pair, with direct consequences for how fungicide-triage deployments should be certified.