Today's digest
How it works
A second observer role is trained inside the same model to answer verifiable questions about the assistant's actual behavior, not just its outputs.