Text-derived Bayesian reconstruction of causal state transitions in railway perimeter safety events

Railway perimeter safety events are rarely determined at the moment an external disturbance enters the operating boundary; their consequences emerge through subsequent state transitions that are recorded only incompletely in accident narratives. This study develops a text-derived Bayesian state-transition framework that reconstructs these narratives as auditable probabilistic state flows. Grounded coding defines the causal states, sentence-level coding identifies directed transitions, and first- and second-order Bayesian models are combined with absorbing-state analysis to characterize local continuation, path dependence and eventual outcome convergence. Among 335 screened cases, 281 contributed transition evidence to a 90-state model. The results reveal that path dependence is selective rather than universal. Preceding mechanisms can strongly redirect the successor distribution of the same intermediate state, yet adding preceding-state information reduces out-of-sample performance when the corresponding triadic evidence is absent and improves it only when that path is sufficiently represented. More historical context is therefore not inherently more informative under sparse narrative evidence. A complementary pattern emerges between evolutionary distance and outcome certainty: upstream and mid-chain states retain longer remaining paths but unstable terminal destinations, whereas near-outcome states have short remaining paths and highly stable outcome convergence. The framework therefore shifts perimeter-safety analysis from identifying hazardous factors to determining when history changes the future of a state and when that future has become sufficiently constrained for targeted intervention.