Coverage-driven alignment - What ‘Teaching Claude Why’ can borrow from AV verification
Proposes borrowing coverage-driven verification from autonomous-vehicle testing to systematically train and evaluate alignment, using explicit coverage maps over scenario dimensions for long-horizon AI agents; largely a conceptual proposal without empirical results.