Big models keep getting better at benchmarks, but reliability under messy real world inputs still feels stuck in place.