The third is competence: people must be able to check the process. When you speak to ChatGPT or Claude, you cannot really inspect the process. Chain-of-thought is hidden from most users. The result is effectively: trust us. With an open model running on a machine you trust, you can inspect how it reflects on its context.