Evaluation of Causal Attention-Sensitivity Audits for Artificial Intelligence Agriculture Pest-Diagnosis Models Under Field-Image Distribution Shift: An Interpretability Robustness Bound for Smallholder Spray-Advisory Reliability

Abstract

Pest-diagnosis models deployed in smallholder spray-advisory systems are audited for interpretability robustness, yet the statistical behaviour of attention-sensitivity audits under field-image distribution shift remains uncharacterised. the present analysis formalises the causal attention-sensitivity audit as a before-after-control-impact design and derives an interpretability robustness bound that links audit power to the intraclass correlation of attention maps across field conditions. Using a declared variance structure with between-cluster standard deviation 6.0, residual standard deviation 4.5 and intraclass correlation 0.64, the operating characteristics of the audit were 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.54 clusters are required to detect a 5 percentage-point shift at 80 per cent power, and that the minimum detectable effect at 80 per cent power declines steeply as cluster count rises from 8 to 32. The bound implies that spray-advisory reliability claims rest on audit designs whose power is frequently inadequate at realistic cluster counts, and that interpretability audits must be sized against the intraclass correlation of attention maps rather than against image-level sample size alone.