Benchmarking Evaluation of Causal Discovery Algorithms for Computer Agriculture Intercropping Yield Synergy Under Unobserved Plot-Level Confounders: An Artificial Intelligence Edge-Orientation Error Bound for Extension Recommendations

Causal discovery algorithms are increasingly proposed as tools for identifying intercropping yield synergy from observational plot data in computer agriculture, yet their outputs are typically evaluated against computed ground-truth graphs that bear little resemblance to the confounding structures encountered in real smallholder systems. The framework is formalised through a structural causal model for intercropping yield in which a latent plot fertility variable induces dependence between crop yields and management decisions. A replication-based benchmarking protocol is specified across a design grid of 8, 16, 24 and 32 clusters, effect sizes of 3, 5 and 8 percentage points, and 2000 replications per cell, with a between-cluster standard deviation of 6.0, a residual standard deviation of 4.5 and an intraclass correlation of 0.64. The derived bound indicates that edge-orientation error increases monotonically with the latent confounding strength and that the minimum detectable synergy effect at 80% power is approximately 5 percentage points when 54 clusters are available. The contribution is a design-stage error bound that allows extension services to judge whether a causal discovery output is reliable enough to inform intercropping recommendations, rather than a claim of empirical discovery from field data.