The probability that reactivity feedbacks will fail to prevent damage is computed by propagating data and modeling uncertainties through transient calculations, with these uncertainties being constrained by experimental evidence. Screening processes are used to identify the most important parameters and accident initiators. The notion of treating an accident initiator in a probabilistic manner is introduced. The response surface method is used to facilitate the error propagation, and a Monte Carlo rejection technique is used to force the parameter variations to be consistent with the observed distribution of experimental quantities. The reliability of the failure probability estimates is evaluated. This method is illustrated by analyzing anticipated transients without scram for the Experimental Breeder Reactor II. The rod run-in initiator is represented by using a reactivity insertion magnitude distribution, a much less threatening and more realistic description than the technical specification limit on rod worths. Reactivity feedbacks are shown to reduce damage frequencies by orders of magnitude, and the experimental constraints are found to have a large effect.