The Favorite That Fell

Start with the miss, because it's the story.

Ilia Topuria walked onto the White House lawn a massive favorite and our headline pick — and Justin Gaethje knocked him out in the fourth round. A near-5-to-1 favorite, the lightweight champion, finished. The biggest upset the sport has produced in some time, and we had it wrong.

But here's the honest part, and it matters more than the L. Topuria carried a HIGH-confidence label at 77.4% into the night — and even that oversold what the blended engine actually thought. The 4-input stack had Topuria at just 56.9%, its lowest tier (LOW), a hair above a coin flip. Every individual voice liked him; the blended engine did not commit, and we've written before that this exact gap — confident-looking label, uncertain underlying number — is where favorites get you. A 56.9% pick losing isn't a model failure. It's a coin landing tails on a single flip. The lesson isn't 'the model blew it'; it's that the HIGH tier promised more certainty than the stack's 56.9% ever did. We'll be tightening how we show that.

The Upset We Did Call

While the main event burned the favorites, the co-main rewarded the contrarians — including us. We picked Ciryl Gane over Alex Pereira, a low-confidence lean against the betting line, and flagged it in our preview as a genuine model/market divergence. Gane knocked Pereira out in the second round, taking the interim heavyweight title and ending — for now — Pereira's bid to become the first three-division champion in UFC history. When you make a point of a divergence going in, you get to point at it coming out. This one we got right.

A Night Without a Single Decision

The stat that defined the card: all seven main-card fights ended by knockout or TKO. Not one reached the scorecards. Gane stopped Pereira, Sean O'Malley stopped Aiemann Zahabi, Josh Hokit stopped Derrick Lewis, Mauricio Ruffy flattened Michael Chandler in the first, Bo Nickal finished Kyle Daukaus in the first, and Diego Lopes stopped Steve Garcia. A uniformly violent night on the sport's biggest-ever stage.

How the Model Did

On winners, the engine went 6-of-7 — strong, but the one it missed was the one everyone was watching, so we lead with that, not the hit rate. Method was the harder night: with every fight ending in a knockout, the model's decision-leaning method calls were always going to struggle. For context, the production model runs about 70% accuracy on the 2025-2026 leak-free holdout (0.6986, n=428); one card — especially one marquee upset — moves nothing about that baseline. We don't put a single-card ROI number on these; our betting grading lives with the Vegas desk, not here.

The takeaway isn't the scoreboard. It's that the model was quietly uncertain on the fight it 'locked,' and quietly contrarian on the fight it called — and on this night, both of those honest signals were the right ones to trust.