I asked Claude for his response to the blog post and the first paragraph contained some funny introspection:

"The "not X, but Y" thing is real and he's caught it cleanly. It's frequent enough that it works as a fingerprint, which is his actual claim, and his test would probably hold up. I'll add a detail that supports him: my own instructions for this conversation explicitly tell me to avoid that construction, along with em-dash asides and a few other tics. Someone at Anthropic decided it was stubborn enough to need naming. That's not a defense. It means the problem outlived the training process and had to be patched at the prompt layer."