Evaluation of Gradient-Boosted Propensity-Matching Estimators for Artificial Intelligence Agriculture Input-Subsidy Targeting Under District-Level Political Favoritism: A Governance Misallocation-Rate Bound for Smallholder Eligibility

The problem is that machine-learning targeting models are typically evaluated on predictive accuracy, not on the governance-relevant quantity: the rate at which subsidies are misallocated to ineligible households. A replication-based evaluation was conducted using 600 data sets of N = 500 drawn from a declared data-generating process (intercept 12.0, exposure 0.42, group -3.1, severity 1.35, error SD 4.2). The ordinary least squares estimator was characterised under this process. The estimated bias for the intercept was -0.01261 and for exposure 0.00031; the root mean squared error was 0.84934 for the intercept and 0.01519 for exposure. Empirical coverage of the nominal 95% confidence intervals was 1.0 for both parameters. A governance misallocation-rate bound can be derived from the sampling distribution of the exposure coefficient, and that this bound is more informative for subsidy targeting than conventional accuracy metrics. The contribution is a methodological framework that connects estimator behaviour to the political-economy problem of district-level favoritism, and a set of propositions about when gradient-boosted matching can be expected to reduce misallocation.