A distribution-free conformal prediction certification layer for safety-critical ordinal classification. Guarantees contiguous prediction sets, marginal/group-conditional validity, true-label coverage under banded reporting noise, and severity-weighted risk control (fatal-omission bounds) with an explicit additive slack budget.
Beta. Statistically honest analysis of road hazards and near misses for safe-streets advocacy: every published rate carries its confidence interval and its denominator, because a count without exposure is not a risk. The committed city datasets are synthetic demonstrations; the reviewed 2020-2024 NHTSA FARS reference surface is real.
