Inverse-Reward Evaluation for Computer Agriculture Harvest-Timing Advice Under Unobserved Farmer Risk Preferences: An Artificial Intelligence Policy-Recoverability Bound

Harvest-timing advice systems in computer agriculture must be evaluated against the risk preferences of the farmers they serve, yet those preferences are rarely observed and cannot be elicited at scale. The framework is formalised as a constrained inverse-reward problem in which a risk-aversion coefficient is identified from a finite set of harvest-timing choices, and a recoverability bound is derived that links the Fisher information of the choice model to the worst-case error in evaluating any candidate advice policy. The bound is then applied to a real country-year panel of agricultural and infrastructural indicators drawn from the World Bank World Development Indicators for 24 African countries over 2000 to 2023 (N = 560), estimated by ordinary least squares with heteroskedasticity-robust standard errors. The estimated model accounts for 59.1 per cent of the variance in GDP per capita (constant 2015 US$), with access to electricity and urban population share entering positively and a time trend entering negatively. These estimates are used to calibrate the scale of the recoverability problem rather than to identify farmer preferences directly. The recoverability bound tightens sharply when the observed choice set is small, when the risk-preference parameter is near risk neutrality, and when the advice policy is sensitive to that parameter. The contribution is a formal criterion that tells an evaluator when inverse-reward evaluation of harvest-timing advice is informative and when it is not, together with a calibration that makes the criterion operational for African agricultural settings.