Post-hoc Conformal Risk Control for Artificial Intelligence Flood Early Warning in Agriculture Under Upstream Gauge Failure: A Computer Safety Guarantee Evaluation

Flood early warning systems that serve agricultural users depend on upstream river gauges, and when those gauges fail the warning model's inputs shift in ways that its training distribution never anticipated. The construction treats gauge failure as a covariate shift on the input space, calibrates a nonconformity score on a held-out reference sample, and applies a risk-controlling quantile to the calibration scores so that the resulting sets satisfy a distribution-free guarantee under exchangeability of calibration and test points within each failure stratum. A formal result shows that the coverage guarantee degrades gracefully as the fraction of failed gauges grows, with the degradation bounded by the total variation distance between the failure-stratified and the pooled calibration distributions. The evaluation is a design and power analysis of the before-after-control-impact estimator used to audit the risk-control layer in a prospective agricultural deployment; no observational data are used or claimed. Under a declared variance structure with between-cluster standard deviation 6.0, residual standard deviation 4.5 and intraclass correlation 0.64, 54 clusters are required to detect a 5 percentage-point reduction in missed-flood rate at 80% power, and 2000 replications per design cell characterise the operating characteristics across cluster counts of 8, 16, 24 and 32 and effect sizes of 3, 5 and 8 percentage points. The contribution is a safety-guarantee formulation that is compatible with existing operational flood-warning pipelines and an audit design that can detect whether the guarantee holds in the field.