Isotonic Post-Hoc Recalibration of Artificial Intelligence Agriculture Disease Classifiers Under Deployment Prevalence Drift: A Computer Reliability-Diagram Evaluation

Deployed plant-disease classifiers are trained on balanced or convenience-sampled data but operate under prevalence regimes that differ from the training distribution, so their nominal probabilities cease to be calibrated and threshold-based decisions inherit a systematic bias. the present analysis treats post-hoc isotonic recalibration as a reliability-diagram problem and asks what a before-after-control-impact design can establish about its operating characteristics when prevalence drifts between deployment waves. No observational data are used or claimed: the analysis characterises the design under a declared variance structure with between-cluster SD = 6.0, residual SD = 4.5 and intraclass correlation = 0.640, evaluated across 2000 replications per design cell for cluster counts of 8, 16, 24 and 32 and effects of 3, 5 and 8 percentage points. The design requires 52 clusters to detect a 5-percentage-point effect at 80% power. The conclusion is that isotonic recalibration is defensible as a post-hoc correction only when the reliability diagram is estimated on deployment-distribution data and the evaluation design is powered for the drift magnitude actually anticipated.