Evaluation of Integrated Gradients Stability Audits for Computer Vision Agriculture Weed-Detection Models Under Herbicide-Resistance Regime Shifts: An Interpretability Attribution-Drift Bound for Smallholder Misclassification Liability

Weed-detection models deployed on smallholder farms are audited for attribution stability, but the audit itself is rarely evaluated as a design. the present analysis treats Integrated Gradients stability auditing as an estimation problem and derives a finite-sample bound on attribution drift under herbicide-resistance regime shifts. The bound separates three sources of drift: baseline-path sensitivity, regime-induced feature-covariance shift, and finite replication error in the attribution estimate. A before-after-control-impact design is characterised analytically under a declared variance structure (between-cluster SD = 6.0, residual SD = 4.5, intraclass correlation = 0.64), with power estimated 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 design requires 54 clusters to detect a 5-percentage-point drift at 80% power. The analysis is methodological: no observational data are used or claimed. The contribution is a bound that converts an interpretability audit into a liability-relevant quantity, and a design characterisation that tells auditors how many clusters a stability audit needs before its verdict can be defended.