Ordinal Misclassification Cost Learning for Artificial Intelligence Agriculture Disease Triage Under Asymmetric Smallholder Loss Functions: A Computer Decision-Regret Evaluation

Ordinal disease-triage classifiers for smallholder agriculture are typically trained under symmetric loss, yet the economic cost of a missed severe infection is not the cost of a false alarm. the present analysis formalises the triage problem as cost-sensitive ordinal classification and derives a decision-regret bound for the asymmetric loss regime that characterises smallholder disease management. The analysis treats the loss matrix as a design object rather than a nuisance: it shows that the regret of a plug-in cost-weighted predictor decomposes into an ordinal-threshold term and a cost-ratio term, and that the two terms trade off against one another as the asymmetry of the loss matrix increases. A replication-based evaluation of the before-after-control-impact estimator was conducted 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. The design achieved 80% power at 54 clusters for a 5 percentage-point effect. Decision-regret, not accuracy, is the correct evaluation target for triage systems deployed where the cost of under-triage is borne by the farmer, and it specifies the conditions under which cost-weighted ordinal learning reduces regret relative to symmetric baselines.