Conformal Prediction Sets for Artificial Intelligence Agriculture Disease Diagnosis Under Class-Prior Shift Across Agro-Ecological Zones: A Computer Coverage-Guarantee Evaluation

Conformal prediction has been proposed as a distribution-free method for equipping agricultural disease classifiers with prediction sets that carry a finite-sample coverage guarantee. The guarantee, however, is marginal over the calibration and test distribution, and it is therefore silent about what happens when a classifier trained in one agro-ecological zone is deployed in another where the class prior differs. the present analysis formalises the coverage gap that arises under class-prior shift, derives a bound on the achievable coverage of split-conformal prediction sets as a function of the total-variation distance between calibration and deployment priors, and evaluates the bound through a replication-based replication study of 600 data sets of N = 500 drawn from a declared data-generating process with intercept 12.0, exposure 0.42, group -3.1, severity 1.35 and error SD 4.2. The estimator behaviour is characterised: the intercept bias was 0.03065 with RMSE 0.90759, the exposure bias was -0.00047 with RMSE 0.01687, and empirical coverage of the nominal 95% intervals was 1.0 for both parameters. Prior shift is not a nuisance to be corrected post hoc but a structural condition that determines which conformal variants can be guaranteed to transfer across zones, and it identifies the conditions under which class-conditional and label-conditional calibration recover the nominal guarantee.