The main event was the one fight our model refused to commit to, and it still landed on the right side.

Mason Jones beat Axel Sola by decision in the third round, the kind of close, distance result that matched the 51% edge our model gave Jones going in. We flagged the bout a TOSS-UP. There was no pretending otherwise. The pick was correct, but on a fight that close, calling it right is closer to a coin landing our way than a forecast we should take much credit for.

The night in brief

Underneath the headliner, the card delivered the finishes the main event withheld. Iwo Baraniewski closed his fight inside the opening round with a knockout. Brando Pericic did the same, ending his bout by first-round KO. Two early stoppages on the same card give the night its shape: a slate that mostly went chalk, with a couple of fighters refusing to let it reach the judges.

The model autopsy

This is the part we publish whether it flatters us or not. On Jones vs Sola, the model graded 11 of 13, an accuracy of 85%. That is a strong night by any honest standard, and it is worth stating plainly before we get to the caveats.

The flagged picks carried it. Our highest-conviction calls, the LOCK and HIGH tier we ask readers to weigh most heavily, went 4 for 4. Every fight we were willing to put real confidence behind came in. The clearest example was Mantas Kondratavicius, a LOCK at 82%, who won as expected. When the model commits, the card rewarded it.

So where did the two losses come from? Not from the picks we underlined. Both misses sat in the unflagged, lower-conviction portion of the slate, the fights where the model itself signaled less certainty. That is the distribution we want when we get things wrong: the errors landing on the bouts we hedged, not the ones we told readers to trust.

The headliner deserves its own line in the ledger because it is the easiest place to oversell a result. We picked Jones. Jones won. But we graded the fight a TOSS-UP at 51%, and a correct call on a near-even split is not evidence the model "read" the fight. It is evidence the model knew it could not. We will take the point in the standings and decline the bragging rights.

No LOCK missed. No HIGH-conviction pick missed. The headliner went our way. On this card, the failures were confined to exactly the places the model flagged as uncertain. That is the failure mode you design for, and the only kind worth shrugging off.

One honest takeaway

A single 85% card does not move what we know about the model over the long run, and we are not going to let it. Our canonical live 2026 record sits at 67.9% across all picks, and 81.4% on the LOCK and HIGH calls we flag for added weight (n=70). Jones vs Sola is one data point that landed above the all-picks line and roughly on the flagged line. That is the right way to read it: a good night that is consistent with the flagged-pick discipline holding up, not proof of anything new.

The pattern that matters is the one we keep returning to. When the model speaks up, it has been worth listening to more often than not. When it stays quiet, as it did on a third-round decision between two fighters separated by a single percentage point, it is telling you something too. On a night when 11 of 13 came in and every flagged pick held, the most useful sentence we can write is the one about the fight we never claimed to know.