"a strange, bizarre form of awareness that is combined with an inability to recognize tells slightly out of distribution" seems like a pretty good description of the models' epistemics in many non-verifiable domains, including some that are presumably being specifically trained for. (For example, Opus clearly has some ability to detect when I am frustrated and write relevant memory files in response, but the "lessons" it records are often bizarrely rigid, broadly scoped prohibitions that do more harm than good.) I think it is exactly what you would expect early signs of eval awareness to look like.