Evaluation of Counterfactual Latent-Shift Audits for Artificial Intelligence Agriculture Yield Models Under Irrigation-Expansion Regimes: An Interpretability Extrapolation Bound for Smallholder Credit Scoring

Smallholder credit scoring increasingly relies on yield models trained in rainfed districts and then applied to households entering newly irrigated schemes, where the joint distribution of covariates and yields shifts in ways that standard interpretability diagnostics do not detect. The audit is formalised as a constrained transport problem between the training and deployment latent distributions, and the bound is expressed as a function of the Wasserstein displacement of the latent representation and the local Lipschitz constant of the attribution map. A replication-based evaluation of the before-after-control-impact estimator was conducted to characterise the operating characteristics of the audit 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 for cluster counts of 8, 16, 24 and 32 and effects of 3, 5 and 8 percentage points. The design required 52 clusters to detect a 5 percentage point effect at 80% power. The bound implies that attribution stability, not attribution magnitude, is the quantity that credit committees should monitor when scoring households under irrigation expansion.