“Any problem domain with formally verifiable solutions and a trusted reward signal can be fully automated with current LLM-based technology.”