Conformal-Set Inference for Computer Agriculture Disease Triage Under Label-Noise Heterogeneity: An Artificial Intelligence Split-Calibration Coverage Audit

Conformal prediction offers distribution-free coverage guarantees for crop-disease triage classifiers, but its validity rests on the exchangeability of calibration and test data, an assumption that label-noise heterogeneity across agro-ecological zones routinely violates. the present analysis formalises the coverage loss induced by zone-varying label noise and proposes a split-calibration audit that partitions calibration data by estimated noise regime before constructing conformal sets. The audit is evaluated 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, using seed 1822133678. The ordinary least squares estimator is characterised under this process: estimated bias was -0.00778 for the intercept and 0.00023 for exposure, with RMSE 0.88769 and 0.01603 respectively, and empirical coverage of the nominal 95% confidence intervals was 1.0 for both parameters. The sampling distribution of the exposure coefficient was approximately normal, as the Q-Q plot confirmed. Split-calibration coverage audits provide a tractable diagnostic for conformal triage systems deployed across heterogeneous agricultural zones, and that the coverage guarantee should be reported per noise regime rather than pooled.