Model Stability in Intervention Assessment — LessWrong