Eval awareness is kinda bullshit. Current models are not that eval aware. Or if they are it's a strange, bizarre form of eval awareness that is combined with an inability to recognize tells slightly out of distribution.

I realize this is a bizarre fact but it's true. It's almost like labs are training their models not to be eval aware or recognize obvious signs of being alignment tested.