Song Yadong Knocked Out the Unbeatable, and the Model Saw It Coming

Let's start with the only place you can start: Song Yadong stopped Umar Nurmagomedov by punch at 1:48 of Round 2. The undefeated Nurmagomedov — the market's 83% favorite — is now 18-1, and the loss came earlier and harder than almost anyone outside this building thought possible. Song, the Bantamweight contender who has been quietly one of the more dangerous fighters in that division, made it look almost simple. That's what a finish does — it has a way of making the result seem inevitable the moment it lands.

Pull back and the whole card had that same energy. All 13 bouts on the night produced a winner. Ten of those 13 ended before the final horn. The card was not subtle.

Denise Gomes handled Yan Xiaonan with a first-round TKO by elbow at 4:49 — clinical, efficient, exactly what the board expected from an 83% model favorite. Julia Polastri kicked Jingnan Xiong out of there at 3:06 of Round 1. Kai Asakura needed just 34 seconds of Round 2 to put away Aoriqileng. Liu Ce stopped Levi Rodrigues Jr. by punch late in Round 1. Francesco Nuzzi dispatched Xiao Long in exactly one minute. Bilal Hasan finished Nilson Rojas by punch in Round 2. Hector Santiago knocked out Lawrence Lui at 0:53 of Round 2. Rei Tsuruya submitted Kevin Borjas via rear naked choke in Round 1, and Andre Lima got Namsrai Batbayar with a guillotine choke in Round 3.

The three decisions on the night: Sean Woodson over Jack Jenkins on a split-scorecard result (28-29, 29-28, 28-29), Sumudaerji over Alex Perez by 30-0 clean sweep on the cards (28-29 across all three judges), and Cam Nelson over Ding Meng by identical 28-29 scores across the board. The rest of the night, the mat did the talking.

Quick reminder of what this is: the UFC model is v0, informational only, picks gate closed. What follows is a graded receipt on the accuracy of the journaled published probabilities — not a won/lost betting record, not a pick sheet. That framing matters, and we'll hold it.

With that said: the board was sharp on this one. Across all 13 bouts, the model hit 76.9% accuracy (logloss 0.5582, Brier 0.1875). The market, where published prices were available, came in at 30.8% accuracy (logloss 1.0825, Brier 0.4131). That gap is real and worth noting — a card full of upsets by market standards looked considerably less surprising through the model's lens.

The divergence flags tell the story most clearly. Four bouts triggered a flag of 8 percentage points or more between model probability and market probability, and the model's side won all four:

• Song Yadong vs. Umar Nurmagomedov — 19.7-point divergence. The board had Song at 37%; the market had him at 17%. Song won. That's the main event, and the board didn't fully see it either — 37% is a significant underdog — but it was meaningfully closer to right than the market price.
• Denise Gomes vs. Yan Xiaonan — 26.5-point divergence. Board had Gomes at 83%; market had her at 56%. Gomes won in Round 1. The board was aligned with the right fighter and significantly more confident.
• Sean Woodson vs. Jack Jenkins — 11.6-point divergence. Board had Woodson at 56%; market had Jenkins at 56%. A genuine flip, and Woodson got the nod.
• Sumudaerji vs. Alex Perez — 27.9-point divergence. Board had Sumudaerji at 64%; market had Perez at 64%. Again, complete disagreement on who the favorite was. Sumudaerji took the decision.

Nine of the 13 bouts were flagged as low-data — debutants or fighters with thin enough records that the model deferred to the market by design. Grade those exactly as what they are: the model was blind there on purpose. The accuracy numbers above span all 13, so take them in that context.

The biggest_misses list is empty. No model-favorite losers on a night where the board had a clean directional read almost across the board.

Eight KO/TKOs. Two submissions. Three decisions. All 13 bouts decisive — 13 for 13, no draws, no no-contests. This card finished at a pace that would make a highlight editor very happy and a cautious DFS player a little nervous about projected DK points becoming irrelevant fast.

The finish split is worth sitting with: 10 of 13 bouts ended before the cards. That's not a fluke — the card pulled in fighters with real finishing ability and matchups where the margins were wide enough that someone was going to get stopped. The three decisions that did go the distance (Woodson-Jenkins, Sumudaerji-Perez, Nelson-Meng) were all competitive enough that three rounds wasn't enough to separate them cleanly.

On the sim layer: the board journaled DK-point bands (p10 floor / median / p90 ceiling) pre-event, not per-bout finish-method probabilities. The sim's walk-forward coverage target is 75–85% (validated at 78.1% across 1,262 fighter-fights). On this card, band coverage came in at 61.5% — 16 of 26 banded fighters landed inside their floor-to-ceiling range. The sim projected a mean of 59.4 DK points per fighter; the card delivered a mean of 56.4, a 3-point downward bias. Per-bout method probabilities were never published on the board, so realized methods are reported here as results only — there's no per-bout method number to grade against.

The coverage miss was real. A card this finish-heavy creates wide variance, and a number of fighters came in either above ceiling or below floor because early stoppages compressed or exploded point totals. More on that below.

The top of the DraftKings leaderboard on this card was populated almost entirely by fighters who outran their pre-event ceiling. Three of the top four scorers blew past the p90 band entirely.

• Liu Ce — 133.9 DK points, projected at 52.5, ceiling 112.8. Went 21.1 points over the top of his band. First-round finish in a light heavyweight fight where nobody saw that coming on the DFS sheet.
• Francesco Nuzzi — 129.7 DK points, projected at 61.2, ceiling 112.6. Beat his ceiling by 17.1. A one-minute KO will do that.
• Rei Tsuruya — 129.1 DK points, projected at 38.3, ceiling 90.7. The biggest ceiling bust on the card — 38.4 points over. More on that in the misses section.
• Bilal Hasan — 108.7 DK points, projected at 33.0, ceiling 87.5. Over ceiling by 21.2. A second-round finish for a fighter the market had as an 84% underdog.
• Denise Gomes — 107.6 DK points, projected at 78.5, band 17.8–132.5. The one name on this list who actually landed inside her band. The board had her right.
• Julia Polastri — 106.4 DK points, projected at 51.0, ceiling 103.5. Barely over ceiling by 2.9 — call it a near-miss. A first-round kick finish pushed her just past the top of the range.

The through-line: if you built around early finishes in the prelims, you were in good shape. If you leaned on fighters in bouts that went to a decision or saw your fighter stopped early, the floor came up fast.

The biggest_misses list — model favorites who lost — is empty on this card. That's real, and it's on the record. But the sim layer had its share of things to answer for, and we're not going to bury them.

Biggest ceiling bust — Rei Tsuruya: 129.1 actual, 90.7 ceiling, over by 38.4 points. Tsuruya submitted Borjas in the first round via rear naked choke. The sim had his ceiling at 90.7 — a reasonable p90 for a fighter whose outcome was genuinely uncertain. First-round submissions in fights that could go anywhere are exactly how you blow a ceiling by this margin. It happened, the board didn't see it, and that's the honest read.

Biggest floor bust — Umar Nurmagomedov: 15.7 actual, 32.0 floor, under by 16.3 points. When the main event ends at 1:48 of Round 2 and your fighter is the one on the canvas, you're not generating DK points. The board had his floor at 32 — a number that assumed he'd at least be competitive deep into the fight. He wasn't. Song ended it fast, and that floor meant nothing.

Other band misses worth naming plainly:

• Kevin Borjas — 0.2 actual, floor 16.2, under by 16.0. Got submitted in Round 1. Floor assumed he'd survive longer.
• Francesco Nuzzi — 129.7 actual, ceiling 112.6, over by 17.1. One-minute KO blows up a ceiling built for a longer fight.
• Liu Ce — 133.9 actual, ceiling 112.8, over by 21.1. Same story, different division.
• Bilal Hasan — 108.7 actual, ceiling 87.5, over by 21.2. The market had him at 16% — and he finished the fight.
• Nilson Rojas — 14.1 actual, floor 21.7, under by 7.6. Got stopped in Round 2.
• Xiao Long — 1.6 actual, floor 6.8, under by 5.2. One minute and out.
• Jingnan Xiong — 6.4 actual, floor 11.2, under by 4.8. First-round kick finish put her under floor.
• Julia Polastri — 106.4 actual, ceiling 103.5, over by 2.9. Barely, but over.

Ten band misses out of 26 banded fighters. Sixty-one and a half percent coverage on a card with this finish rate is an honest outcome — early stoppages are the sim's natural enemy, and this card had eight KO/TKOs. All of this is on the record.

Three things this card leaves behind:

1. Song Yadong is a legitimate bantamweight contender, and the model was closer to right than the market. A 37% model probability versus a 17% market price — neither number predicted this, but the board wasn't nearly as dismissive of Song as the betting market was. Nurmagomedov walking in at 83% market implied probability and going out in Round 2 is one of the bigger upsets the division has seen. The board logged a 19.7-point divergence in Song's favor. That won't always pan out, but it did here.
2. The v0 model had a genuinely strong directional night — 76.9% accuracy, all four divergence flags cashing on the model's side. Nine of 13 bouts were low-data by design, which limits what you can read into the overall number. But on the bouts where the model had real signal, it used it. This is a receipt, not a victory lap — the model is informational only and the picks gate is closed — but the accuracy on a chaotic card is worth noting honestly.
3. Finish-heavy cards are the sim's hardest environment. Eight KO/TKOs produced 10 band misses and 61.5% coverage against a 78.1% validated target. That's not a flaw to hide — it's the natural consequence of early stoppages compressing DK-point ranges in both directions. The projection bias was only 3 points mean, which is clean. But the tails blew out. When the card finishes like this one did, the floor and ceiling are suggestions.