They can spot evaluation awareness because it appears in reasoning tokens.

It is still helpful I believe, but your point is well taken. The problem is that we have relatively few tools for monitoring alignment, and longer loops of processing that stay in latent space means less ability to monitor.