Evaluation of Byzantine-Resilient Federated Learning for Artificial Intelligence Agriculture Extension Consortia Under Ministry-Level Model Poisoning: A Governance Convergence Bound for Advisory Integrity

Ministry-level actors in agricultural extension consortia can poison the shared model that federated learning is meant to protect, and the governance question is not whether robust aggregation works in the abstract but what integrity guarantee a consortium can credibly promise its smallholders. The bound is derived analytically from the geometry of the robust aggregation rule and the political constraint that no single ministry may control more than a fixed share of update weight. Integrity is jointly produced by the statistical robustness of the aggregator and the institutional design of the consortium, and that neither alone suffices. Three governance levers follow: a participation cap calibrated to the aggregator's breakdown point, a public audit trigger tied to the bound, and a convergence criterion that treats advisory drift as a political as well as a statistical event. The contribution is a bridge between the machine-learning literature on Byzantine robustness and the political science literature on delegated authority in extension systems.