What could go wrong: adults and children calibrate predictions and explanations of others' actions based on relative reward and danger