Half the Bracket Is Dead. Michigan 79, Duke 78 Is Not.

By Collin Lee | March 28, 2026

With 0.7 seconds left in the West regional semifinal, our bracket was about to lose another limb. Texas had just tied the game on an and-one layup by Dailyn Swain. Purdue brought the ball up. Braden Smith drove the lane. His layup bounced off the backboard. Off the rim. And then Trey Kaufman-Renn - who had been invisible on the boards all game - appeared from nowhere, tipped the ball in, and kept our West bracket alive by less than a second.

Purdue 79, Texas 77. Our bracket had Purdue advancing at 75% confidence. It took a controversial tip-in with 0.7 seconds remaining to cash that ticket.

That's the Sweet 16 in miniature. Nothing came easy. Every game was a story. And when the dust settled, the bracket looks like a house after a storm - half the structure is gone, but the two load-bearing walls are still standing.

5 out of 8. Same hit rate as the Round of 32. Same three Final Four picks alive - Duke, Arizona, Michigan. Same one missing - Florida, whose absence has now rippled through the entire South region. And the same championship prediction sitting at the center of it all, unmoved: Michigan 79, Duke 78.

That prediction has now survived three rounds, 56 games, and 16 incorrect picks. It was there before the tournament. It was there after the R64. It was there after Florida went down. And it's here now, unchanged, while half the bracket around it crumbles.

The full results

Here's every original bracket Sweet 16 prediction against what actually happened. Asterisks mark where the actual opponent was different from what we predicted.

* Predicted opponent was eliminated in a prior round; our pick faced a different team.

Five correct, three wrong. One dead pick from a prior round (Florida). Two teams that were there and lost - Houston and Iowa State. Here's the breakdown:

• Florida (dead) - Still gone. Still hurting. The South remains a wasteland.
• Houston at 72% - Illinois held Houston to 34% shooting in Houston's own building and won 65-55. Two double-doubles from the Illini bigs. Houston had no answer. A genuine upset and a real model miss.
• Iowa State at 64% - Tennessee physically dominated Iowa State without their All-American Joshua Jefferson (ankle injury). The Vols outrebounded the Cyclones 43-22 and pounded them inside 42-22 in points in the paint. Our model can't account for injuries, and Jefferson's absence was devastating.

What went right

The East is still perfect

I keep waiting for the other shoe to drop in the East, and it keeps not dropping.

Duke 80, St. John's 75. The Red Storm held a double-digit second-half lead. Duke stormed back, powered by Isaiah Evans's 25 points on 10-for-15 shooting and a crucial step-back three. Dylan Darling - the man who hit the buzzer-beater against Kansas to get St. John's here - pulled up for a deep contested three with 4 seconds left, trailing by three. It missed wide left. Bracket saved.

UConn 67, Michigan State 63. UConn jumped out to a 25-6 lead, nearly blew a 19-point cushion, then closed the door with championship-caliber free throw shooting down the stretch. Tarris Reed Jr. hit 4 of 4 at the line in the final minute despite shooting 59% from the stripe on the season. Alex Karaban added a dagger three with 1:39 left.

Through three full rounds, every team the East bracket picked to advance has advanced. The only miss in the entire region - Ohio State over TCU in the Round of 64 - didn't matter, because Duke handled TCU even more easily than it would have handled Ohio State. Every R32 pick correct. Every Sweet 16 pick correct. And the Elite 8 matchup in the East is (1) Duke vs. (2) UConn - exactly, precisely, what our bracket said before a single game was played.

One harmless miss in 14 games. In a sport defined by madness, the East has been pure order.

The West holds too

Arizona put up 109 points against Arkansas - 63.8% from the field, six players in double figures. Purdue survived by a fingertip against Texas. Different vibes, same result: both teams advanced, and the West Elite 8 matchup is (1) Arizona vs. (2) Purdue - also exactly what our bracket predicted.

Two of our four Elite 8 matchups are perfect. Not adjusted, not rerouted. The pre-tournament bracket nailed the East E8 and the West E8 through three rounds of chaos. The model saw Duke, UConn, Arizona, and Purdue as the class of their regions, and all four delivered.

Michigan looks like a champion

Michigan 90, Alabama 77. Yaxel Lendeborg had 23 points, 12 rebounds, and 7 assists - one assist short of a triple-double. Michigan trailed at halftime, then outscored Alabama 43-28 in the second half. They shot 48% from three. Their bench outscored Alabama's reserves 33-6.

Our championship pick is playing like a championship team. The model gave Michigan a 93% chance in the Sweet 16 (against the originally predicted Texas Tech) and Michigan won by 13 against a tougher opponent. Whatever you think about the model, Michigan is making it look smart.

What went wrong

The South is gone. All of it.

Florida in the Round of 32. Houston in the Sweet 16. The two teams our bracket had contesting the South regional final are both on planes home. Instead, 9-seed Iowa will face 3-seed Illinois for a spot in the Final Four. Our original bracket has no opinion on this game because neither team was supposed to be here.

The Houston loss is the one that stings. Our model had Houston at 72% over Illinois - a moderate but genuine confidence. Houston was playing in their own building, at Toyota Center in Houston. They were last year's national runner-up. And Illinois held them to 22-of-64 from the field, a 34% shooting performance that no model could have predicted. David Mirkovic and Keaton Wagler combined for 27 points and 22 rebounds. The Illini simply suffocated Houston's offense.

This is the kind of result that makes you question whether the model adequately accounts for defensive matchup specifics. Illinois's length and rebounding created problems that don't show up as a single number in a feature vector. The model sees "Houston has a top-10 efficiency rating." It doesn't see "Illinois's front line is going to grab 16 offensive rebounds and make Houston's life miserable for 40 minutes."

Iowa State and the injury problem

Our model had Iowa State at 64% over Virginia in the original bracket. Virginia was out, replaced by Tennessee. Tennessee won 76-62, and it wasn't that close.

The headline reason: Iowa State's All-American Joshua Jefferson was out with an ankle injury. Tennessee outrebounded the Cyclones 43-22 and dominated the paint. Without Jefferson, Iowa State couldn't match Tennessee's physicality.

This is a known limitation we've talked about since the first post. The model can't account for injuries announced the morning of a game. Jefferson's absence turned a competitive matchup into a mismatch. At 64% confidence, this was already one of our weakest Sweet 16 picks. With an injury, it became untenable.

The damage report

Our original bracket predicted 8 teams in the Elite 8. Here's how reality compares:

5 out of 8. The pattern is now unmistakable: the East and West are pristine, the South is destroyed, and the Midwest lost its second seed. Three of the four regions have at least one correct E8 pick. The South has zero.

But look at what survived. Four of the Big Ten's teams made the Elite 8 - Michigan, Purdue, Illinois, and Iowa. And two of our four Elite 8 matchups are exactly as predicted: Duke vs. UConn in the East and Arizona vs. Purdue in the West. The bracket isn't just getting the teams right - it's getting the games right.

What's still alive

Our champion pick, Michigan, just outscored Alabama 43-28 in the second half to win 90-77. Lendeborg is playing at an All-American level. The bench is 33-6 in outscoring opponents' reserves. Michigan set a program record for wins this season. They look every bit the team the model thinks they are.

Our runner-up pick, Duke, survived a St. John's comeback and won 80-75. Not pretty, not dominant, but they survived. Caleb Foster returned from a foot fracture and scored 11 second-half points. The model doesn't know Foster is back. That's a variable working in Duke's favor that isn't in any feature vector.

Our West Final Four pick, Arizona, dropped 109 on Arkansas. The most dominant Sweet 16 performance in the tournament. The model had Arizona at 93%. Arizona played like 99%.

Three of four Final Four picks alive. Same three since the Round of 32. Same championship prediction. Same score.

What the model says about the Elite 8

I almost didn't bother this time. I already knew what it would say. But I fed the actual Elite 8 field into the same model anyway. Two of these matchups are exactly what the original bracket predicted. Two are new. Here's the path:

What changed (and what, remarkably, didn't)

Four times now. Before the tournament, after the R64, after the R32, and again after the Sweet 16. Same teams. Same score. Same one-point margin. The model has watched 56 games go by, 16 of them incorrectly predicted, an entire region of the bracket reduced to rubble, and its championship prediction has not changed by a single point.

I don't know if that's conviction or stubbornness. At some point the distinction stops mattering.

What did change:

• The South E8 is Iowa vs. Illinois (92% Illinois). Neither team was in our original bracket's South regional plans. The model strongly favors Illinois, and if Illinois wins, they face Duke in the Final Four. The model gives Duke a 98% chance in that game. So the South bracket's complete destruction might paradoxically give Duke an easier path to the championship than the original bracket predicted.
• Michigan vs. Tennessee (91%) replaces Michigan vs. Iowa State (87%). The model sees Tennessee as slightly less threatening than Iowa State was. Michigan's path to the Final Four actually got marginally easier.
• Arizona vs. Purdue (81%) and Duke vs. UConn (93%) are unchanged. These are the original bracket's predictions, word for word. No adjustments needed. The model called these matchups before the tournament started and three rounds later, here they are.

The part that's hard to say out loud

Here's the thing I've been avoiding all series.

I'm a Miami guy. ACC. And the model's runner-up - the team that's supposed to carry my bracket to the championship game - is Duke. I've been rooting for Duke for three weeks. I've been needing Duke to win for my bracket to work. Every time they survive, my bracket survives. Every time Cameron Boozer scores, my championship prediction stays alive.

The system was built to remove emotion from the bracket. It succeeded. What it couldn't remove was the emotion of needing a team I've spent my whole life rooting against. That's the cruelest joke the algorithm has played on me, and it's been playing it since game one.

When St. John's had that double-digit lead on Duke in the second half, some part of me - the part that grew up hating Duke - felt a flicker of something. Not quite hope. Just... familiarity. The comfortable feeling of watching Duke lose. And then Duke came back, and Evans hit that step-back three, and I pumped my fist. For Duke. Because the bracket needed Duke. Because Michigan 79, Duke 78 needed Duke to get there first.

Removing emotion from the bracket didn't make me neutral. It made me root for things I never would have chosen. The algorithm doesn't know I hate Duke. It just knows Duke is really, really good at basketball. And now, three rounds in, I need them to be good for two more games.

The things you do for a prediction.

What's next

The Elite 8 starts Saturday. Here are the games that matter most:

• Duke vs. UConn (93%) - The game our bracket predicted before the season. Duke needs to win for the championship prediction to survive. UConn blew a 19-point lead against Michigan State and survived - are they rattled or battle-tested? The model says Duke at 93%. If you'd told me before the tournament I'd be hoping for a Duke win, I wouldn't have believed you.
• Arizona vs. Purdue (81%) - The other E8 matchup our original bracket nailed. Arizona just scored 109. Purdue needed a controversial tip-in at the buzzer. Momentum says Arizona. The model agrees at 81%.
• Michigan vs. Tennessee (91%) - Our championship pick against the team that just physically dominated Iowa State. Tennessee grabbed 43 rebounds and 16 offensive boards against the Cyclones. If they bring that energy against Michigan, the model's 91% confidence will be tested. Lendeborg's near-triple-double says Michigan is ready. The Vols' glass-eating says this won't be easy.

If Duke, Arizona, and Michigan all win, both Final Four matchups will be exactly what the original bracket predicted: Duke vs. the South winner in one semi, Michigan vs. Arizona in the other. We'd be two games from the exact championship game - Michigan vs. Duke - that the model has been calling since before a single ball was tipped.

Four times, same prediction. Michigan 79, Duke 78. The machine still doesn't flinch. I still can't stop watching.

Update: The Elite 8 is done. Duke is out. The prediction is dead. But the champion pick isn't.