Evaluation of Constrained Multi-Objective Reinforcement Learning for Computer Agriculture Land-Use Zoning Under Competing Tenure Claims: An Artificial Intelligence Pareto-Frontier Audit for Expropriation Risk

Agricultural land-use zoning in sub-Saharan Africa must reconcile competing tenure claims with productivity and equity objectives, yet the algorithms increasingly proposed for this task are evaluated almost exclusively on their ability to identify Pareto-efficient allocations. Pareto-frontier optimality is a necessary but insufficient criterion for zoning systems that allocate land under contested customary and statutory tenure, because a frontier can be efficient while systematically concentrating expropriation risk on politically marginal claimants. The framework is formalised as a constrained Markov decision process with a lexicographic constraint structure, and the audit is specified as a set of frontier-level diagnostics: constraint-violation density, risk concentration across tenure classes, and frontier stability under preference perturbation. No empirical data are analysed; the contribution is methodological and conceptual. The choice of constraint architecture, not the choice of learning algorithm, determines whether a zoning policy is tenure-safe, and that audits reporting only hypervolume or sparsity cannot distinguish tenure-safe from tenure-unsafe frontiers. The proposed audit provides a reproducible diagnostic that regulators can apply to any multi-objective zoning model before deployment.