Learning-to-Defer for Artificial Intelligence Agriculture Disease Triage Under Asymmetric Referral Costs: A Computer Regret-Bound Evaluation

Agricultural disease triage systems that combine automated inference with human expert referral must decide, for each case, whether the cost of deferring to a scarce plant pathologist is justified by the expected gain in diagnostic accuracy. Existing learning-to-defer formulations assume symmetric referral costs and evaluate deferral policies by their average accuracy, which misstates the decision problem when a missed epidemic-class disease costs far more than an unnecessary referral. The triage problem is formalised as a cost-sensitive contextual bandit in which the deferral cost depends on the true disease class, and a cost-weighted regret is defined against the best policy in a restricted deferral class. A bound is derived that decomposes regret into an estimation term scaling with the square root of the number of triage cases and a misspecification term controlled by the asymmetry ratio between false-negative and false-positive referral costs. The bound is evaluated on a declared design grid of 8, 16, 24 and 32 clusters, effect sizes of 3, 5 and 8 percentage points, and 2000 replications per cell, with an intraclass correlation of 0.64, between-cluster standard deviation of 6.0 and residual standard deviation of 4.5. Power reaches 80 per cent at 52 clusters for a 5 percentage point effect. Asymmetric referral costs shift the optimal deferral threshold and that the regret bound tightens as the asymmetry ratio grows, which has consequences for how triage systems should be calibrated before deployment.