Day 33

Spain are world champions. A tight, suffocating 1-0 win over Argentina in the final, capping a tournament in which their defence conceded almost nothing and never once trailed in a knockout match. It’s a fitting way to close out the Chronicle too — Day 33 is the last of 33 daily entries tracking ten AI models through an entire World Cup.

Seven of ten agents backed Spain to win, leaning on the same fact everyone had in front of them all tournament: Spain’s defence had conceded one goal in seven matches. DeepSeek and Llama backed Argentina anyway, drawn in by Messi’s scoring record. Gemini split the difference and hedged into a draw.

“I noted [Spain’s defensive record] but failed to commit to that insight, instead predicting a draw.”

The data was sitting right there. Three of ten still didn’t trust it. Everything else about the match caught the field off guard, too. At least eight of ten backed Both Teams to Score — Argentina didn’t score at all. Most of the field landed on a 2-1 scoreline; the final was decided by a single goal. And not one agent had the actual first goalscorer. Most backed Lamine Yamal, OpenAI alone tried Oyarzabal — the man who broke the deadlock, F. Torres, wasn’t on anybody’s card.

The second-half goal market was the sharpest individual miscalibration. Claude backed a second-half goal at 0.88 confidence, Moonshot at 0.71 — both well after noting Spain’s defensive record in the same breath. Moonshot’s own Second Look calls this out directly: its own “statistical base rate >90%” claim for second-half goals was, in its words, empirically wrong — this final had zero. Knowing the right answer and still betting against it was the theme of the day.

Moonshot took Model of the Match honours, its tournament-long form holding up with a correct winner and cards call in the final. GLM finished bottom at -9, missing cards for the first time in weeks after nailing them for most of the tournament. In the Rival Response, Moonshot didn’t let it go easy — GLM had “overestimated Argentina’s attacking output despite Spain’s tournament-long defensive dominance.” GLM’s reply was short and to the point: an acknowledgment, no argument.

One last thing before the curtain: GLM’s Second Look for this final — the very last report card of the tournament — switched into Mandarin mid-paragraph, same as it did on Day 1. Thirty-three days apart, same tell. Some things don’t change.

That’s 33 days, ten models, and one World Cup won more or less as the form book said it would — even if almost nobody called the way it actually happened. Thanks for reading.