Benchmarking Evaluation of Retrieval-Augmented Agronomic Advisory Systems for Computer Agriculture Under Stale Extension Corpora: An Artificial Intelligence Citation-Faithfulness Audit for Smallholder Safety

Retrieval-augmented advisory systems are increasingly proposed as a way to deliver agronomic guidance to smallholder farmers, but their safety depends on whether the citations they attach to recommendations actually support those recommendations when the underlying extension corpus is stale. The framework defines a citation-faithfulness score over claim-evidence pairs, specifies a stale-corpus stress protocol, and derives the operating characteristics of the audit under a declared variance structure. A replication-based evaluation of the before-after-control-impact estimator was conducted across 2000 design cells with between-cluster SD = 6.0, residual SD = 4.5 and intraclass correlation = 0.640; 52 clusters were required to detect a 5 percentage-point effect at 80% power. The analysis characterises how audit power varies with cluster count and effect size, and identifies the minimum detectable effect at 80% power for cluster counts of 8, 16, 24 and 32. The contribution is a reproducible audit design that separates retrieval failure from generation failure in citation-faithfulness benchmarking, together with the design parameters needed to power such audits in smallholder advisory contexts.