The public record

How grading works

A pick is frozen when it is published, before the card. After the event, each frozen pick is compared with the official result and marked a win or a loss on the winner call alone — the method call (how the fight ended) is graded separately and never flatters the headline. Nothing is regraded, reworded or withdrawn because it aged badly; the only rows that ever leave the table are ones whose source export is corrupt, and when that happens the page says which card and how many rows.

Confidence tiers, what they are tested against, and the places the method is weak are set out on the methodology page, which is longer than this one and does not flatter us.

This page grades the direction of each pick. Whether the confidence number beside it is honest — when we publish 70%, do seven in ten of those picks win — is measured on the calibration page, year by year, off the same graded rows.

One change inside the 2026 season. The served model changed on 5 September 2026: the number beside each pick was a single voice's probability up to and including the 29 August 2026 card, and the four-voice stack from the 5 September 2026 card on. From the 3 October 2026 card, a bout involving a fighter making a UFC debut is published with the market's own pick instead, with no tier; those 4 picks count in every total here and in no tier row. Direction is graded here the same way either side of that line; the calibration page shows the two as separate series, because one curve across both would describe a model this site no longer serves.