Notes on control evaluations for safety cases — LessWrong