Causal Fairness Auditing for Artificial Intelligence Agriculture Credit Allocation Under Historical Land-Tenure Bias: A Computer Counterfactual-Robustness Evaluation

Agricultural credit algorithms trained on historical lending records inherit the land-tenure exclusions embedded in those records, yet conventional fairness audits evaluate only observed group disparities and therefore cannot distinguish a model that reproduces tenure bias from one that merely reflects it. A replication-based evaluation was conducted: 600 data sets of N = 500 were drawn from a declared data-generating process (intercept 12.0, exposure 0.42, group -3.1, severity 1.35, error SD 4.2), and the sampling behaviour of the estimator was characterised. Across the 600 replications the intercept bias was -0.04737 with RMSE 0.83244, the exposure bias was 0.00061 with RMSE 0.01504, and empirical coverage of the nominal 95% intervals was 1.0 for both parameters. The results indicate that the estimator is approximately unbiased for the exposure coefficient under the declared process, that its interval coverage is conservative rather than nominal, and that the intercept is the parameter most exposed to finite-sample distortion. Counterfactual robustness, not observed parity, should anchor the audit, and it specifies the conditions under which a credit-allocation model can be certified as tenure-robust.