Selective-Ensemble Evaluation for Computer Agriculture Drought-Advisory Models Under Regime-Switching Rainfall Forecasts: An Artificial Intelligence Regret-Bound Audit

Drought-advisory models in computer agriculture are typically evaluated by average predictive accuracy, yet the operational cost of an advisory depends on which model is selected when rainfall regimes shift. The audit treats model selection as a sequential decision problem in which a meta-learner observes forecast features and chooses among candidate advisories, and it bounds the excess loss of the selective ensemble relative to the best fixed advisory in hindsight. Using NASA POWER monthly reanalysis for Lagos, Nigeria (6.52°N, 3.38°E), 2000–2023, 312 monthly observations, the study establishes the empirical regime structure that any such audit must respect: mean air temperature at 2 m was 26.414 °C (SD 0.943 °C) and the warming trend was 0.1553 °C per decade. Formal results give a high-probability regret bound for the selective ensemble that decomposes into a switching penalty, an estimation penalty and a misspecification penalty, and the bound is shown to be tight when the regime process is persistent. The audit reframes evaluation from average accuracy to worst-regime regret, and it identifies the conditions under which a selective ensemble is preferable to the best single advisory.