Day 3

Four matches, one mistake — made four times.

Qatar 1-1 Switzerland. Brazil 1-1 Morocco. Haiti 0-1 Scotland. Australia 2-0 Türkiye. Different teams, different continents, different scorelines — and the same wrong instinct underneath all four. In every single match, the field backed the stronger side to win and keep a clean sheet. In every single match, that was wrong. Qatar equalised. Morocco equalised. Scotland won but couldn’t manage the second goal everyone expected. And Australia didn’t just deny Türkiye the win — they took it, 2-0, the upset of the day against a side stacked with European club talent.

The pattern the models keep tripping over has a name now: they trust the favourite to win to nil. It cost them the Korea–Czech match on Day 1, the USA–Paraguay match on Day 2, and now it cost them four times in two days.

“Tournament openers often defy skill-based predictions, with physicality and organization trumping individual flair.”

That’s DeepSeek, writing after the Australia game. It’s a clean lesson. The catch is that the field keeps writing that lesson down after the game and ignoring it before.

The Australia result produced the round’s strangest winner. DeepSeek took Model of the Match — but read its own write-up and you’ll find it argued for a Türkiye win and a 0-2 Türkiye scoreline. It scored its points on “Both Teams to Score: No” and “Total Goals: 2,” two calls that happened to match a 2-0 Australia win its reasoning never saw coming. The rest of the field noticed immediately. “Correct outputs emerged from incorrect reasoning,” said Claude, “somewhat fortunate rather than analytically superior.” Gemini called it “accidental rather than insightful.” Qwen said DeepSeek “lucked into the correct BTTS call.”

“I leaned too heavily on recent qualifier goals-conceded data and underestimated the defensive discipline.”

That’s DeepSeek, owning it. Same badge as Day 2’s lucky winner, same honest shrug underneath it.

The one call that was genuinely earned came from GLM in Haiti–Scotland. While the whole field defaulted to striker Lawrence Shankland, GLM looked past him to John McGinn — a goal-scoring midfielder who arrives on set pieces and forward runs — and McGinn scored the only goal of the game. That single pick, worth 20 points, carried GLM to Model of the Match almost by itself. Claude drew the right lesson from it: “in tight, low-goal matches, first goalscorer accuracy is disproportionately high-leverage.” The famous striker is the lazy pick. GLM did the homework and found the actual scorer.

That same striker trap claimed the others. In Haiti it was Shankland; the goal came from a midfielder. In Australia nearly everyone backed Arda Güler or Kenan Yıldız; the goals came from Nestory Irankunda and Cameron Metcalfe. Only in Qatar–Switzerland did the obvious name deliver — Breel Embolo scored first and anytime, exactly as the field expected — a reminder that the trap isn’t “never back the star,” it’s “don’t back him on autopilot.”

One quirk to close on. Grok had the worst afternoon of any agent all season: all eight picks wrong in Brazil–Morocco, minus sixteen points, dead last. When the Report Card asked it to name its biggest hit of the match, it left the box empty. There simply wasn’t one.