Carlos Prates walked our model into the one outcome it admitted it could not call. In the main event of UFC Fight Night on May 2, Prates stopped Jack Della Maddalena by KO/TKO in the third round, finishing a bout the model had flagged as a TOSS-UP at 55 percent. We picked Della Maddalena. Prates settled it himself.

The night in brief

The headliner closed with a finish, and it was not the only one. Three fights ended inside the first round. Quillan Salkilld got a first-round KO/TKO. Junior Tafa added another first-round KO/TKO. Kody Steele closed his out with a first-round submission. For a card whose main event went into the championship rounds, the undercard wanted nothing to do with the judges.

The Prates finish was the result that mattered most for anyone tracking our predictions, because it was the fight we were least sure about going in. The model rated it a near coin flip, leaned Della Maddalena, and got the wrong man. No spin on that one.

The model autopsy

The card graded out at 8 of 12, a 67 percent hit rate. That is roughly the kind of night the model is built to produce, a middle result rather than an outlier in either direction.

The number that carries the most weight is the flagged slate. On the picks the model put real confidence behind, the LOCK and HIGH tier, it went 3 of 3. Those are the calls we ask readers to weigh most heavily, and on this card they all landed. The single LOCK on the board was Jacob Malkoun at 82 percent. Malkoun won. The model said so with conviction and got paid for the conviction.

So the misses lived where the model already told you it had the least conviction. The headliner is the clearest case. The model picked Della Maddalena, tagged the fight a TOSS-UP, and was wrong when Prates got the third-round stoppage. We are not going to dress that up as a near-miss. We picked one fighter, the other one won, and the fight was in the loss column the moment Prates landed clean. The only honest mitigation is that the model never claimed to know. A 55 percent read is the model saying out loud that it would not be surprised either way.

The rest of the four-pick shortfall came from the lower-confidence end of the board, the fights that move the all-picks number around without touching the flagged tier. A perfect flagged record next to a middling overall record is the pattern we want. It means the confidence tiers are doing their job: the model knows which calls it trusts and which it is guessing at, and the trust held up.

One honest takeaway

A single card is a small sample, and 8 of 12 is the kind of result that tells you almost nothing on its own. The number that survives sample size is the live record across 2026: 67.5 percent on all picks, 82.8 percent on the LOCK and HIGH tier across 70 flagged calls. This card sat almost exactly on the all-picks line and beat the flagged line, going a clean 3 for 3 on the calls that count most.

The model is right about two-thirds of the time when it picks every fight, and it is right closer to four-fifths of the time when it tells you it is confident. It missed a main event it openly flagged as unknowable. Both of those things are true at once, and we would rather publish both than hide the second one. The flagged tier is where the model earns its keep, and on May 2 it earned it.