Evaluation of Temporal-Conformal Prediction for Computer Agriculture Pest-Risk Forecasting Under Phenology-Stage Shift: An Artificial Intelligence Coverage-Guarantee Audit for Insecticide-Timing Safety

Temporal-conformal predictors are increasingly proposed as a way to attach distribution-free coverage guarantees to pest-risk forecasts that inform insecticide timing, yet the guarantee is marginal over the calibration window and can fail when the phenology stage of the crop shifts between calibration and deployment. the present analysis treats the coverage guarantee as an auditable object rather than a property to be assumed. It formalises phenology-stage shift as a covariate shift on the stage-conditional risk distribution, derives the conditions under which split-conformal coverage degrades, and proposes a stage-stratified audit that reports coverage per phenology stage rather than in aggregate. The audit is demonstrated on a replication-based evaluation of a linear risk-score estimator with intercept, exposure, group and severity terms, using 600 replications of N = 500 observations drawn from a declared data-generating process (intercept 12.0, exposure 0.42, group -3.1, severity 1.35, error SD 4.2). The estimator was characterised by ordinary least squares. Estimated bias was -0.01322 for the intercept and 0.00028 for exposure; root-mean-square error was 0.90474 and 0.01575 respectively; empirical coverage of the nominal 95% intervals was 1.0 for both parameters. The audit shows that aggregate coverage can be satisfied while stage-conditional coverage is not, and that the gap is a function of the stage-shift magnitude and the calibration-stage composition. The contribution is a coverage-guarantee audit protocol and a formal statement of when temporal-conformal pest-risk forecasts remain safe for insecticide-timing decisions.