The FightIQ model published exactly one LOCK at UFC Vegas 119, the loudest call on the board: Andre Lima at 82.7% confidence. He lost a unanimous decision. That makes two straight cards where the pick we put the most weight behind at the top got burned, and the night's clean-looking 75% hides exactly why.

The model went 9 for 12 on winner picks at UFC Vegas 119 in Las Vegas on June 21, a 75% night. That is the wrong number to lead with. Of the 12 fights, the model called the right method on only 6, and all three of its winner misses came in the same place: fights that went to the judges. On the seven bouts that ended by finish, the model was a perfect 7 for 7. On the five that went to a decision, it went 2 for 5.

The honest framing, then, is not the 75%. The model reads finishes well and decisions poorly, and on this card the conviction sat on the decisions.

The lead: we lost the one pick we flagged hardest

The model published exactly one LOCK at Vegas 119. It was Andre Lima, rated at 82.7% confidence in the id6 flyweight bout against Kevin Borjas. Lima lost a unanimous decision. The single most confident call on the board was the only LOCK, and it missed.

That stings more in context. One week earlier, at Freedom 250 on June 15, the model's loudest pick was a HIGH on Ilia Topuria at 77.4% confidence. Topuria was knocked out. So across two straight cards, the model's flagged top-of-card pick has been burned back to back.

Be precise about what that sentence does and does not say. Topuria was a HIGH, not a LOCK. Lima was a LOCK. They were different confidence tiers, and the pattern is not \"the LOCK lost twice.\" The pattern is narrower and still uncomfortable: on consecutive events, the pick the model put the most weight behind at the top of the card was wrong.

The calibration story, and what we are doing about it

This is the part that matters most, and we will state it as an observation rather than a diagnosis.

The confidence tier we publish, the LOCK and HIGH labels, comes from one layer of the system, the V14 model. Behind that label sits a blended four-model stack that produces its own probability. On both of the burned picks, the blend was far more cautious than the published label.

On Topuria, V14 said HIGH at 77.4%. The four-model stack said 56.9%, only a LOW. On Lima, V14 said LOCK at 82.7%. The stack said 58%, again only a LOW. On both cards the cautious blend was the better read, and the confident single-model label was not.

We will not dress that up. The published tier ran hot on the two picks where it cost the most, and the quieter consensus number would have told a more honest story both times. A calibration review to align the published confidence tier with the stack is in flight with our CTO and quality-control review. It is work in progress, not a shipped fix, and we will grade it in public the same way we are grading this card.

What the model got right

The misses are the headline, but they are not the whole night.

The model went 7 for 7 on every fight that ended inside the distance. KO/TKO finishes: 4 for 4. Submissions: 3 for 3. When a fight resolved by a finish, the model was on the right fighter every single time.

Four of the six flagged picks, the LOCK and HIGH tier combined, hit. The four HIGH winners: Navajo Stirling stayed unbeaten, Vinicius Oliveira got his finish, Bia Mesquita closed with a first-round armbar, and Murtazali Magomedov won the night's $100,000 twister submission. That last one is the kind of result the model is built to like, a clear favorite finishing in the manner the method engine leaned toward.

And the headliner. The stack rated Manel Kape a LOW-confidence pick at 56.3% over Kyoji Horiguchi, a near-even read that barely tilted his way. Kape stopped Horiguchi by KO in the third round. We will not pretend that was a forecast cashing. A 56% call landing is a coin flip going our way, and we will report it as exactly that.

The three misses, named

For the record, all three winner misses went the distance:

The LOCK, Andre Lima at 82.7%, lost a unanimous decision to Kevin Borjas. In the middle of the card, where the model leaned Michael Aswell, the pick lost as Gaston Bolanos took a unanimous decision. And the LOW lean on Allan Nascimento lost a split decision to Mitch Raposo, a genuine coin flip.

Three misses, three decisions, zero finishes missed. That is a consistent signal, and it is the signal the calibration work is chasing.

The takeaway

The number FightIQ tracks is the high-conviction one, not the all-picks accuracy, because the high-conviction number is the one that would matter if anyone leaned on it. Our established canon stands: a 67.9% live all-picks record across 70 fights in 2026, 81.4% on LOCK and HIGH picks, against a leak-free backtest holdout of roughly 70% all-picks and 84.3% on LOCKs. Our backtest team is regenerating the season LOCK-and-HIGH tally to fold in this card, which added a flagged miss to that column.

We publish this autopsy because the credibility is the product. A model that only shows you its hits is a marketing deck. This one went 9 for 12, nailed every finish, called the headliner coin flip right, and lost the single pick it flagged loudest about for the second card in a row. All of that is true at the same time, and you get to see all of it.