Evaluation of Federated-Learning Aggregation Audits for Artificial Intelligence Agriculture Disease-Model Consortia Under Ministry Parameter-Poisoning: A Governance Contamination-Rate Bound for Smallholder Outbreak Detection

Federated learning allows agricultural ministries and model consortia to train crop-disease classifiers across districts without centralising smallholder field data, but the aggregation server remains a point of political vulnerability: a ministry with write access to the parameter stream can poison the global model while leaving local training untouched. The work is a formal and methodological analysis, not an empirical evaluation. It specifies a contamination-rate bound as a function of audit coverage, audit sensitivity, the number of participating districts and the adversary's share of aggregation rounds, and it characterises the detection power of the design under a declared variance structure (between-cluster SD = 6.0, residual SD = 4.5, intraclass correlation = 0.64) using a before-after-control-impact estimator across 2000 replications per design cell. The intraclass correlation of 0.64 substantially inflates the cluster requirement: 54 clusters are needed to detect a 5 percentage-point effect at 80% power, a figure that exceeds the 8, 16, 24 and 32 cluster counts examined and that therefore constrains what a realistic consortium can audit. Aggregation audits should be understood as a governance instrument with a computable contamination ceiling rather than as a technical add-on, and it derives the conditions under which the ceiling is tight enough to protect smallholder outbreak detection.