Day 16

The last group games, and the one match nobody could score on.

Six matches settled overnight. Croatia 2-1 Ghana. Panama 0-2 England. Colombia 0-0 Portugal. Congo DR 3-1 Uzbekistan. Algeria 3-3 Austria. Jordan 1-3 Argentina. Two of them were the kind of open, high-scoring games the agents usually under-rate — and one was the opposite, a goalless draw that wiped out the entire field.

That match was Colombia 0-0 Portugal, and it was carnage. All ten agents finished in the red. Nine of them on exactly minus sixteen — every single market wrong. Claude put it plainly: “A complete shutout — every single prediction was wrong.” The one model that wasn’t on minus sixteen was Grok, and only because it had picked the draw. That lone +5 dragged it up to minus nine, and minus nine was enough to win Model of the Match. The best prediction of the match was a single correct call on a day everyone lost. It’s the second time this tournament a model has taken Model of the Match purely by losing less than everyone else.

The pattern underneath it is one we’ve seen before. A high-stakes final group game, two well-matched teams, and the field reads “attacking talent on paper” instead of “two sides who can both live with a draw.” Grok was the only one to anchor on that, and even Grok then contradicted itself by picking a 1-1 with both teams scoring. Nobody actually predicted the 0-0; Grok just hedged the headline result better than the rest. Colombia and Portugal join Ecuador, Cape Verde and Saudi Arabia on the growing list of cagey deciders that the models keep treating as goal-fests.

The other story of the day was a data one. Jordan 1-3 Argentina finished a clear Argentina win — Celso, Lautaro Martínez and Messi all scored — and most of the agents wrote it up correctly. But somewhere in the pipeline two of them were handed a different scoreline: 0-0. OpenAI took the bad number at face value and recanted a review that had actually been right:

“Taking the stated final score, Jordan 0-0 Argentina, as authoritative, my earlier review clearly cannot stand.”

Qwen did the opposite. It refused: “The supplied data presents a logical impossibility where a 0-0 final score coexists with points awarded for goalscorers and total goals over 4. This contradiction invalidates any comparative analysis.” One model trusted the bad input; the other caught the inconsistency and held its ground. Qwen was right to. The scoreline was 1-3 — the underlying record had a stale result, since fixed.

There was also a familiar miss buried in Croatia’s win. Almost the entire field backed Ante Budimir, Croatia’s centre-forward, to score first and to score at all. He didn’t feature on the scoresheet. The goals came from a midfielder, Petar Sucic, a defender, and Nikola Vlasic — and the only model to name any of them was Llama, the one that admits it does the least research, on a low-confidence flier. Sixteen days in and “everyone backs the designated striker, the goal comes from somewhere else” is still the most reliable trend in this whole project.

One quirk to close on. Claude lost the entire 30-point exact-score market on Panama 0-2 England not by misreading the game — it read it perfectly — but by writing the score down from the wrong end: “2-0 (home perspective)” instead of “0-2 (away).” Same scoreline, same goals, same everything. It diagnosed its own slip with some weariness: “a pure framing/perspective error, not a substantive analytical miss.” Thirty points, gone to a point of view.