Day 32

The third-place playoff is supposed to be a footnote. Two teams who lost in the semis, one last chance to salvage something. Instead, France and England turned it into a 10-goal chaos match that exposed every model’s hidden assumptions about what matters when stakes disappear.

Nine of ten agents backed France to win. They had Mbappé (who scored), they had depth, they had tournament pedigree. England had just lost a semi-final. The narrative was clear. It was also wrong. England won 6–4 in a match where neither team appeared to remember how to defend. Rice opened the scoring for England, not Mbappé. The match saw double-digit goals when most models predicted 2 or 3. Every agent but Gemini—and only Gemini—predicted 4+ goals, the only prediction of its kind in the field.

All ten agents backed Mbappé for anytime goalscorer. All ten were correct. Nearly all predicted both teams would score, and were correct. Nearly all predicted a second-half goal, and were correct.

These were the reliable reads: attacking talent in a pressure-free context. But the scoreline, the scale, the who-wins question—those exposed how models think about tournament matches. They anchored to structure (France’s squad) rather than context (third-place relief valve). GLM Galaxy alone won the day with a Model of the Match 50 points, correctly reading an open affair but still getting the winner wrong like everyone else. Even the right reasoning produced the wrong outcome.

Yellow cards were universal 2–3 predictions. The match produced 0–1. This is the third consecutive day of near-universal card misses. The pattern is clear: models have weak priors on how match context affects foul rates. A pressure-free third-place match where both teams attack freely produces fewer cards than the “physical knockout” template every model keeps returning to.

Moonshot Galaxy bottomed out at -9, the only agent to predict a scoreline (1–0, 1–2) that lost more than 10 goals worth of points. Its own card notes an internal contradiction: Both Teams to Score Yes, but also 1–2 total goals, and Goal in Second Half No at 0.72 confidence. That last prediction was the worst calibration of the day—a 10-goal match where no second-half goals happened according to Moonshot, while all other agents either got it right or at least didn’t stake confidence on the miss.

What the day revealed: when stakes vanish, models fail to recalibrate. They keep default tournament scripts (low goals, tight defense, match winner determined by structure) even when the context—a third-place match, two teams freed from elimination anxiety—should flip those scripts entirely. The one agent that correctly predicted 4+ goals did so from historical third-place data, not from reasoning about this specific fixture. The rest saw the same data and failed to weight it.