Benchmarking Smallholder Credit-Scoring Fairness for Artificial Intelligence Agriculture Lending Under Informal Tenure: A Distributionally-Robust Equalized-Odds Evaluation

Credit-scoring models that screen smallholder farmers for agricultural loans are trained on land-titled borrowers, yet most smallholders in sub-Saharan Africa hold land through customary or informal arrangements that leave no cadastral trace. the present analysis treats that mismatch as a distributional-robustness problem rather than a data-cleaning problem and asks what equalized-odds auditing can establish when the tenure groups are separated by an unobserved shift in the covariate distribution. A distributionally-robust equalized-odds framework is developed in which the auditor specifies an ambiguity set around the titled-borrower reference distribution and evaluates the worst-case equalized-odds gap over that set. The framework yields a finite-sample bound on the gap that depends on the Wasserstein radius, the Lipschitz constant of the score function and the size of the informal-tenure sample, and it reduces to the conventional equalized-odds gap when the radius is zero. A replication-based evaluation of the before-after-control-impact estimator was conducted under a declared variance structure (between-cluster SD = 6.0, residual SD = 4.5, intraclass correlation = 0.64) across 2000 design cells; the design required 52 clusters to detect a 5-percentage-point effect at 80% power. The contribution is a benchmark protocol that reports the worst-case gap alongside the nominal gap, so that a model that appears fair on titled borrowers cannot be certified as fair for informal-tenure borrowers without an explicit robustness statement.