Yeah, but it takes a heck of a lot more compute to to run an llm than a regex. So I think there's still some value in obfuscation.

That is a fair observation, especially since my default Claude model is Opus 4.6, which is about as far away from efficient as you can get. I don't have any recent models downloaded on my MBP, but maybe this evening I'll try it again and see how that goes -- especially the smaller lightweight models.