Graph-Neural Emulation of Convective Storm Tracks Under Sparse Rain-Gauge Telemetry: Calibration Bounds and Warning-Lead Equity for Artificial Intelligence Nowcasting

Convective storm nowcasting in regions with sparse rain-gauge telemetry faces a structural tension: graph-neural emulators can interpolate storm tracks from incomplete observations, but the calibration bounds that govern their reliability, and the equity of the warning lead times they produce, remain poorly characterised. The analysis treats the gauge network as a graph with known sampling geometry, defines a spectral calibration bound in terms of the graph Laplacian and the emulator's Lipschitz constant, and shows that lead-time disparities between well-observed and sparsely observed catchments are bounded by the same spectral quantity. A worked illustration using NASA POWER monthly reanalysis for Harare, Zimbabwe (2000–2023, 312 monthly observations, mean 18.914 °C, decadal warming trend 0.2001 °C) anchors the calibration discussion in a real reanalysis record. The framework yields three results: a calibration bound that degrades predictably with gauge density, a lead-time equity condition that fails when spectral gap ratios exceed a threshold, and a set of design constraints for gauge augmentation. The contribution is a bounded, testable account of when graph-neural nowcasting can be trusted for equitable warning, and when it cannot.