Why can’t they have trained the classifier on internal red teaming?
Of course Anthropic realizes saying this straight is problematic so they said they examined request metadata, but no, I don't think they can get this kind of insight from metadata (token counts, request time, etc.)