Model A answers everything, right 24% of the time. Model B declines half and is wrong far less. Which do we ship?

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