Evaluation of Counterfactual Data-Augmentation for Computer Agriculture Rare-Crop Disease Detection Under Herbarium Sampling Bias: An Artificial Intelligence Transportability Bound for Surveillance Equity

Herbarium and specimen collections used to train rare-crop disease classifiers are not random samples of the fields where surveillance is most needed, so a model that performs well on held-out specimens may fail precisely where it matters. Six hundred data sets of N = 500 were drawn from a declared data-generating process (intercept 12.0, exposure 0.42, group -3.1, severity 1.35, error SD 4.2) and the estimator's behaviour was characterised across the replications. The intercept was recovered with estimated bias -0.0015 and RMSE 0.90304, the exposure coefficient with estimated bias 7e-05 and RMSE 0.01614, and empirical coverage of the nominal 95% intervals was 1.0 for both parameters. The bound derived here separates the augmentation's achievable risk reduction from the irreducible component attributable to sampling bias, and shows that the calibration estimator is unbiased and well calibrated under the stated process. The contribution is a formal transportability criterion that lets surveillance programmes decide, before deployment, whether counterfactual augmentation can close a herbarium-to-field gap or merely relocates it.