“According to LangChain's Harrison Chase, many early evaluations at Character AI involved researchers simply 'vibe checking' the systems by interacting with them, rather than using formal metrics.”