I'm not seeing a contradiction. The original claim was that you can't measure robustness without testing live, because you can't reproduce edge cases in a lab. But that's not true; you can reproduce edge cases that are known in a lab. This isn't 100% effective, granted, but it's effective enough that you certainly can test robustness to a reasonable degree. It's like saying that you don't know how safe a car is until you've drive it 100,000 miles on the real roads; Real Life™ will find things you missed in testing, but you can still run enough crashes to get a decent idea of how safe the car is.