AI judgment for production ML. Your monitoring tells you a number crossed a threshold. Arbiter tells you why, whether it matters, and what to do, using open-weight Gemma 4. Suppressed 4 of 5 drift alerts on a measured run, and refuses to retrain on pipeline-corrupted data.