Nearcast-based "deployment problem" analysis — LessWrong