A Trustworthy AI Framework for Nuclear Power Plant Accident Triage With Conformalized Causal Transformers

Abstract

Safety-critical decision-support systems must operate under partial observability and strict time constraints while communicating uncertainty in a statistically defensible manner. In early event identification, the available evidence is often insufficient to uniquely identify the underlying event, making single-label predictions brittle and operationally risky. We present REACT (Robust Event Analysis with Conformal Triage), a framework for sequential decision support that combines a causal Transformer with split conformal prediction to generate calibrated set-valued outputs on a fixed decision-time grid. The causal Transformer processes multivariate telemetry using strict temporal masking, while split conformal calibration converts class scores into prediction sets with finite-sample marginal coverage guarantees under the declared calibration rule. To improve quantile stability when calibration data are sparse at early decision times, REACT supports causal moving-window pooling of nonconformity scores. Under this pooled construction, the formal coverage target is the induced pooled decision-time bin rather than strict time-specific coverage. We further introduce a hierarchical configuration-selection procedure that first enforces validity through lower confidence bounds on empirical coverage and then prioritizes prediction-set efficiency and advisory timeliness. Experiments on the NPPAD benchmark, including ablations of scoring rules, pooling strategies, and advisory thresholds, characterize the empirical trade-offs among coverage, prediction-set size, and advisory timeliness under the evaluated simulator distribution.