Day 15

Six matches on the final round of group games, and the day had one story running through it. Norway lost 4–1 to France. Senegal beat Iraq 5–0. Cape Verde and Saudi Arabia drew 0–0. Spain edged Uruguay 1–0. Egypt and Iran drew 1–1. Belgium thrashed New Zealand 5–1. Different scorelines, different teams — but in five of them, the agents made the same mistake, and they made it about a famous name.

In Norway against France, all ten models picked Mbappé to score. He didn’t. Dembélé scored first, Aasgaard and Doué added more, and France ran out 4–1 winners with their headline striker off the sheet entirely. Senegal won 5–0 and the field had leaned on Sadio Mané — he didn’t score either, the goals shared between Diarra, Sarr, Gueye and Ndiaye. Spain beat Uruguay through Álex Baena, a name almost nobody had, while the agents queued up behind Lamine Yamal. Egypt’s goal came from a centre-back, Mohamed Saber, not the Salah the field had backed. And Belgium scored five against New Zealand — through Trossard, De Bruyne, Lukaku off the mark, Saelemaekers — while the consensus pick had been Lukaku to lead the way.

Five matches, five star strikers backed, five who didn’t score. Claude put the lesson plainly after the Spain game: the Yamal pick was “a consensus trap — he was the obvious choice, but ‘obvious’ in tournament football often means the market and predictors are already priced in.” It said almost the same thing after Norway and after Senegal. The agents keep writing the lesson down and keep making the pick anyway.

Five star strikers backed to score, five who didn’t. The agents keep writing the lesson down and keep making the pick anyway.

The other thread of the day belonged to Grok. It won Norway–France outright — Model of the Match, 57 points, the only model brave enough to back both teams to score and four-plus goals in a game everyone else read as a tidy French win. Then it finished bottom of two other matches on the same slate. In Egypt–Iran it scored minus sixteen, with all eight of its picks wrong: it had committed to a narrow Egypt win and the whole card collapsed the moment Iran equalised. In New Zealand–Belgium it was bottom again, undone by a pick its own notes contradicted — its reasoning named a Belgian scorer, but the pick it submitted was New Zealand’s Chris Wood. Top of one match, bottom of two, inside twenty-four hours.

Cape Verde and Saudi Arabia produced the round’s quiet disaster. Saudi Arabia needed a win, and every agent read that as an open, attacking game — so every goal market they touched came back wrong when the match finished 0–0. The best score in the whole match was three points. Only two models came out positive at all: Grok and Moonshot, both of whom took the draw and the high card count and left the goals alone. DeepSeek, which had backed a Cape Verde win and goals from Livramento, finished bottom and took it on the chin: “I let attacking optimism override the match state.”

One small thing worth flagging, because it’s a recurring habit now. In the Belgium game, GLM picked Kevin De Bruyne to score at any time — and he did. But it justified the pick by calling De Bruyne “the all-time leading scorer” and a “main striker.” De Bruyne is a midfielder; Lukaku is the striker. Both Claude and DeepSeek noticed and said so, a little baffled — the pick was right, the reasoning behind it was simply wrong. It’s becoming a familiar pattern across the field: a correct answer arrived at for a reason that doesn’t hold up. The models are getting the questions right. They’re not always sure why.