Benchmarking Evaluation of Cross-Farm Knowledge Distillation for Computer Agriculture Disease Classifiers Under Seasonal Domain Shift: An Artificial Intelligence Transfer-Risk Bound for Advisory Reliability

Cross-farm deployment of knowledge-distilled disease classifiers is constrained by seasonal domain shift, yet the reliability of advisory outputs under such shift remains unquantified. The bound decomposes the target-farm risk into source-farm empirical risk, a distillation gap term, and a seasonal domain-divergence term, each estimable from unlabelled target-season data. A before-after-control-impact design was used to characterise the estimator's operating characteristics under a declared variance structure (between-cluster SD = 6.0, residual SD = 4.5, intraclass correlation = 0.64), with power estimated across 2000 design cells spanning cluster counts of 8, 16, 24 and 32 and effect sizes of 3, 5 and 8 percentage points. The analysis indicates that 52 clusters are required to detect a 5-percentage-point effect at 80% power, and that the minimum detectable effect declines steeply between 8 and 24 clusters before flattening. The bound provides advisory services with a computable, season-specific reliability certificate rather than a point estimate of accuracy, and the design analysis shows that the certification threshold is attainable at realistic farm-cluster counts only when the distillation gap is held below the seasonal divergence term.