Evaluation of Concept-Probe Transfer Audits for Artificial Intelligence Agriculture Disease Classifiers Under Cross-Regional Deployment: An Interpretability Generalization Bound for Smallholder Fungicide Safety

Concept-probe audits are increasingly proposed as a safeguard for computer-vision crop-disease classifiers deployed across regions, yet the audit's own transfer behaviour is rarely characterised. The audit is modelled as a two-stage procedure: a probe is fitted on source-region activations and then applied to target-region activations to certify that a safety-relevant concept, such as lesion evidence warranting fungicide application, remains decodable. A bound is derived that decomposes the target-region probe error into a source-region empirical term, a concept-distribution divergence term and a finite-sample concentration term. A before-after-control-impact design was used to characterise the audit's operating characteristics under a declared variance structure with between-cluster SD = 6.0, residual SD = 4.5 and intraclass correlation = 0.64; power was estimated across 2000 design cells over cluster counts of 8, 16, 24 and 32 and effect sizes of 3, 5 and 8 percentage points. The analysis indicates that 56 clusters are required for 80% power at a 5 percentage-point effect. The bound implies that audit certification is a statement about probe transferability, not about classifier safety, and that smallholder fungicide decisions require the audit to be paired with an explicit divergence budget.