Our assistant now personalises answers using each user's history. How do you evaluate quality when there's no single right answer?

Make your call first. Use the published answer and rubric to examine your reasoning when you are ready.