Try ObserverBench — from estimates to safer decisions

Test how well an AI's internal monitors catch harmful decisions and edge cases.