Compare base and trained personal models on held-out work tasks, failure cases and privacy boundaries. Use before claiming personalization or enabling serving.
