For a program of interest represented as a sequence of states comprising variables and flags wherein controls transition the program from state to state, and wherein the program includes tests at prescribed locations in the program that are applied upon specific states and generate observables, a method includes inserting at different points in the program hooks that capture program runtime behaviors constituting classification features. The method further includes flattening the program states, commands and tests along a program run-time timeline, and identifying a root cause of a program failure by backtrack analyzing a stream of input states and commands and output consequent states, commands and test results to detect causal correlations between variables, states and commands. The step of backtrack analyzing includes determining joint conditional probabilities and identifying candidate failure root causes that maximize conditional probabilities.