Counterfactual-Policy Evaluation for Computer Agriculture Fertilizer Subsidy Targeting Under Unobserved Plot Fertility: An Artificial Intelligence Off-Policy Regret Bound

Fertilizer subsidy programmes in smallholder agriculture allocate vouchers on observable plot and household characteristics, yet the returns to subsidized fertilizer depend on unobserved plot fertility. the present analysis treats subsidy targeting as an off-policy evaluation problem in which the logging policy is the existing allocation rule and the target policy is a counterfactual allocation rule learned from observational data. Because plot fertility is a hidden confounder, standard inverse-propensity-weighted estimators of policy value are biased, and the sign of the bias is not identified from the observed data alone. The bound is shown to be tight in the sense that it is attained by an adversarial fertility assignment, and a robust targeting rule is proposed that minimizes worst-case regret. The framework is illustrated with a real World Bank panel of 24 African countries observed from 2000 to 2023 (N = 560), in which GDP per capita is regressed on access to electricity and urban population share; the fitted model (R-squared = 0.591, F = 184.764) is used only to calibrate the scale of the confounding parameter, not to estimate subsidy effects. The contribution is a distributionally robust off-policy evaluation bound that makes the fertility-confounding assumption explicit and auditable in computer-agriculture subsidy targeting.