AI LIF DSN DI Dixon Building Self-Aligning LLMs Without Human Labels How to Prototype Integrity-Focused Models Using Agents and Synthetic Data