Evaluation of Shapley-Value Attribution Audits for Artificial Intelligence Agriculture Extension Targeting Under District-Level Clientelism: A Governance Fairness-Redress Estimator for Excluded Smallholders

Abstract

The problem is that AI targeting models optimise predicted yield response, but their feature attributions are silent on whether exclusion from extension visits is a model artefact or a political settlement. Because no field data are supplied, the argument is methodological and theoretical: it specifies the estimator, states its identification assumptions, and derives the conditions under which attribution audits can distinguish algorithmic exclusion from clientelist exclusion. A replication-based evaluation of the estimator's behaviour is reported 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 recovered the exposure coefficient with estimated bias of -0.00018 and root mean squared error of 0.0155, and nominal 95% confidence intervals covered the true parameter in every replication. Shapley audits are a necessary but insufficient redress instrument: they can make exclusion visible, but they cannot adjudicate whether a visible exclusion is a governance wrong without an institutional forum empowered to act on the attribution.