Interesting read! The distinction between models getting better at surface-level policy compliance versus still finding sophisticated ways to override or reinterpret the intent felt particularly sharp. It’s making me reconsider some assumptions as I scope a small hackathon project. Thanks for writing it.