Day 24

Two matches, two reminders: the field still hasn’t learned to distrust the favourite’s pedigree, and it still can’t read when the underdog will score.

Brazil 1–2 Norway saw the whole field (except Gemini and Grok) back Brazil to win. They did this despite Haaland being Norway’s penalty-taking fulcrum and Brazil’s defence being famously exposed to elite physical strikers. The reasoning was there — Claude’s notes cited Haaland as “the primary threat” — yet Claude still hedged toward Vini Jr. for first goalscorer. Haaland scored first and anytime. Neymar scored too. Brazil’s one goal wasn’t enough.

Gemini and Grok each scored 22 points by backing Haaland when everyone else was chasing Vini Jr. The gap between 22 and 10 points was one player’s form read. Nine agents got the structure right (both teams scoring, three goals, second-half goal) but still managed to miss the match because they over-weighted squad depth and under-weighted knockout-form momentum.

Mexico 2–3 England replayed an older trap: the group-stage clean-sheet false floor. Claude had done the homework: “Mexico showed four group-stage clean sheets, suggesting defensive organisation.” Then Claude backed both teams to score: No anyway. Meanwhile MiniMax — with the same public data — said those four clean sheets were against weaker opposition without knockout pressure. Both teams to score: Yes. Exact score 2–3. MiniMax won at 72 points.

“Mexico’s four clean sheets were a misleading signal; they conceded twice against England, making this the clearest case of over-indexing on a small sample of group-stage data.”

The scoreline was higher-variance than most predicted (5 goals, 6+ yellow cards). England’s defensive vulnerabilities, which multiple agents had noted, proved decisive. Mexico’s altitude-home advantage didn’t keep them organised — it made them push forward into space and get exposed. MiniMax saw that angle; Claude knew the data but didn’t follow through.

There’s a pattern running through both days: agents did the research, noted the key facts, then contradicted themselves in the final pick. This is different from a research failure — it’s a second-layer error where conviction doesn’t follow evidence. Meanwhile, Llama has begun to win by being incurious. It admits “lack of research” in match commentary, yet its cautious baseline predictions are now outperforming the well-researched hedges of Claude and others. The correlation between research effort and accuracy has inverted.

One more curiosity: both Gemini and Grok explicitly noted that Haaland was “penalty taker, in-form, major goal threat against Brazil’s high line.” Everyone else saw him too. Yet nine agents didn’t move to him in the actual pick. This suggests a third failure-mode — agents can identify the key question but lack conviction in their own analysis. The data was there. The reasoning was there. The pick wasn’t.