Testing method that compares model outputs against executable ground-truth event sequences, not just final answers.