Techniques include debate, recursive reward modeling, constitutional AI, and using weaker models to flag issues for stronger ones. The problem is circular: the overseer must be trustworthy enough that the smarter system cannot game it. It is a research program, not a single product feature.