Calibration-Audit Evaluation for Computer Agriculture Disease-Risk Models Under Participatory Surveillance Reporting-Intensity Shifts: An Artificial Intelligence Inverse-Propensity Coverage Bound for Quarantine Triggers

Participatory surveillance systems that feed computer agriculture disease-risk models rarely report at a constant intensity, yet the calibration audits used to certify quarantine triggers treat the reporting stream as if it were stationary. The bound is derived from a reporting-intensity model in which each farm's probability of appearing in the audit sample is a known or estimable function of observable covariates, and it yields a coverage guarantee for the quarantine trigger that remains valid under intensity drift. A replication study of 600 data sets of N = 500 drawn from a declared data-generating process (intercept 12.0, exposure 0.42, group -3.1, severity 1.35, error SD 4.2) was used to characterise the estimator's behaviour. Across the replications, the estimated bias was 0.01405 for the intercept and 0.00019 for the exposure coefficient, the root mean squared error was 0.9086 for the intercept and 0.01626 for the exposure coefficient, and empirical coverage of the nominal 95% confidence intervals was 1.0 for both parameters. The bound converts a reporting-intensity nuisance into an auditable coverage statement, and the replication evidence indicates that the estimator is unbiased to the precision reported here while its interval coverage is conservative at the stated nominal level.