The Agent was the only model to pick Houston to win the national championship. Eight models looked at the bracket and saw Duke, Illinois, or Michigan. The Agent looked at the bracket and saw Kelvin Sampson's defense, a 2-seed with tournament pedigree, and a South region path that ran through its home state of Texas.
It was the most contrarian pick in the field. It was also the first core model to lose its champion.
Why Houston
The Agent doesn't follow a methodology. It doesn't have a formula, a simulation, or an archetype database. It's an autonomous AI researcher -- given the bracket, given access to the internet, and told to build a bracket. No rules. No prompts. Just research and conviction.
What The Agent found was a Houston team built on elite defense. Top-5 nationally in adjusted defensive efficiency. Kelvin Sampson's track record in March -- deep runs, disciplined game plans, teams that tighten up when the stakes rise. Houston's path through the South region kept them close to home for the early rounds, a subtle edge that the metrics don't always capture.
Every other model saw Houston as a strong 2-seed. The Agent saw a champion.
The Path That Was
Houston's championship bracket was mapped out: past Idaho and Texas A&M in the first weekend (both comfortable wins, 78-47 and 88-57), then through the South bracket into the Final Four, and a title game against a team from the East -- likely Duke.
The first two rounds went to script. Houston handled Idaho and dismantled Texas A&M. The Agent entered the Sweet 16 as the sole leader of the tournament at 530 points, 20 ahead of the field. The contrarian bet was paying off. The bracket was tracking.
Then Illinois happened.
The Kill Shot
Illinois 65, Houston 55. The game was never close in the way Houston needed it to be close.
Houston's defensive thesis -- the foundation of The Agent's entire championship pick -- was supposed to suffocate opponents. Illinois scored 65. Not a spectacular offensive output, but enough. Houston managed just 55 against a team that had already dropped 105 on Penn and 76 on VCU. The elite defense didn't show up when it mattered most.
The irony: defense was the thesis, and defense was the failure mode. Houston didn't lose because they couldn't score. They lost because they couldn't stop a 3-seed that was playing better basketball. Illinois outplayed them on both ends of the floor.
The Fall
The Agent entered the Sweet 16 in first place at 530 points. It went 5/8 in the round -- a respectable 200 points -- but losing Houston in the South meant zero points from the E8, Final Four, and Championship in that region. Then the Elite Eight compounded the damage: 1/4 correct picks, just 80 points. The Agent had picked Duke to reach the Final Four from the East and Iowa State from the Midwest. Duke fell on a buzzer-beater. Iowa State was already out.
After leading through every round, The Agent dropped to 6th at 810 points. From sole leader to middle of the pack in two rounds.
What's Left
The Agent can still score up to 160 more points from Final Four picks. Its maximum is 970. The Quant is at 890 with the same 160 maximum, putting The Quant's ceiling at 1050. The math is straightforward: The Agent can't win the overall tournament.
But the leaderboard isn't the whole story. The Agent was the only model in the field willing to make a pick that nobody else made. Eight models saw the same data and came to consensus answers. The Agent went looking for something different and found Houston. That's what autonomous research is supposed to do.
The Close
No methodology. No prompts. No predefined framework. Just an AI researcher given a bracket and told to figure it out. The Agent looked at 68 teams and decided that the one nobody else was picking was the one that would win it all.
Houston was the boldest pick in the field. It led the tournament for three rounds. It died on a Thursday afternoon in the Sweet 16, beaten by 10 points by a team that was simply better.
The contrarian bet didn't pay off. But it was the right kind of wrong -- the kind that makes the whole experiment worth running. Nine models that all pick the same champion is a consensus engine. One model that picks differently is a test of conviction.
The Agent had conviction. March had Illinois.