Certified OOD Detection for Artificial Intelligence Climate Emulators Under Unseen Forcing Scenarios: Score Bounds and Seasonal Forecast Safety

Neural climate emulators are increasingly used to produce seasonal forecasts, yet their reliability collapses when forced by radiative or land-surface conditions outside the training distribution. The framework combines a Mahalanobis-style score in a learned latent space with a conformal calibration layer, producing a certified threshold that controls the probability of silently accepting an OOD input. We derive high-probability bounds on the score gap between in-distribution and OOD regimes, and show that the certified threshold is achievable with a calibration sample whose size scales logarithmically in the confidence level. The framework is illustrated on a reanalysis-derived temperature series for Nairobi, Kenya (2000–2023, 312 monthly observations, mean 20.146 °C, warming trend 0.2477 °C per decade), where the emulator's latent score exhibits a clear separation between historical and perturbed forcing regimes. The certified detector attains a bounded false-negative rate under a range of computed forcing perturbations while preserving in-distribution coverage. The analysis clarifies when OOD certification is feasible for climate emulators and when it is not, and it identifies the calibration budget as the binding constraint on seasonal forecast safety.