Uncertainty-Aware Post-Onset Accident Diagnosis: An Audited Within-Cohort Benchmark on NPPAD

We examine point predictions and empirical uncertainty in a reproducible benchmark of 1211 simulated Nuclear Power Plant Accident Data (NPPAD) trajectories from 12 accident classes, including 100 hot-leg loss-of-coolant accident (LOCA) trajectories. Onset is obtained retrospectively from transient reports; the endpoint is post-onset diagnosis, not online event detection. Seven deterministic statistics of 96 shared sensors provide 672 features at 10-, 20-, 30-, 60-, and 96-sample windows. Repeated grouped holdout uses five fixed, stratified partitions, keeping equal complete numerical records in the same training, calibration, or test role. The 1109 equality blocks do not establish independent simulation scenarios; all results are exploratory within-cohort estimates. At 30 samples, histogram gradient boosting and random forest attained mean accuracies of 0.9952 and 0.9927, whereas logistic regression and a multilayer perceptron attained 0.8663 and 0.7908. Random forest LOCA regression yielded a mean absolute error of 0.8206 percentage points and empirical 90% interval coverage of 0.9400 at this window. Exact calibration counts, conformal ranks, thresholds, covered case counts, and binomial reference intervals expose substantial small-sample uncertainty. Five calibration cases per LOCA class force every 90% and 95% Mondrian threshold to infinity. Removing a dimensionally inappropriate ratio and changing numerical scaling reduced historical neural network extremes but did not eliminate physically unreasonable predictions. Sensor corruption and ordered-severity extrapolation degraded clean-calibrated uncertainty. The benchmark supports transparent comparison within this simulator cohort; it does not establish independent-scenario generalization, conformal exchangeability, or plant readiness.