The priority already stored for this email. It is the baseline. It is not one of the models.
Matches existing priority. How often this model chose the same priority already stored for that email.
Catches urgent mail. Of the emails already marked Do today, how many this model also marked Do today. Emails it did not answer count as missed.
Answered clearly. How many emails came back with a real priority, instead of a broken answer.
Typical time. The usual time to answer one email. Half were faster than this, and half were slower.
Cost for these 50. What it cost to answer these emails. Runs on your GPU. No per-email charge.
Blind judges
Most urgent first.
Pick one list, then compare it with another model or with the existing priority. A move up means this list treats the email as more urgent.
Pick emails on the left. Each column is the move that model would have the inbox owner take.
You are the judge.
Read the email and choose what you would do. Model answers stay hidden so they do not sway you. After you save, the scoreboard shows how often each model agrees with you.