A football prediction test shows how to judge Hermes Agents
Before the World Cup group stage began, 12 predicted every group match. Each model received the same prompt and had to give the winning team, exact score, possession, , confidence level, and tactical reasoning. Once a match kicked off, its prediction was locked.
Knockout-stage predictions have not been made yet because the bracket is not known until the group stage ends. After 54 matches, 4.7 led the overall table with 141 points and also had the most exact-score predictions, with 9. Sonar 2 predicted the most match winners, with 35 correct winners.
Grok 4.3 ranked only sixth overall but correctly picked the 10 times, the best result for that category. Scores were based on exact score, correct winner, goal difference, , an underdog bonus, and . One concern remains: not every model was running under exactly the same conditions, but the provided material does not show what those condition differences were.
Key points
- 12 predicted every World Cup group-stage match from the same prompt.
- The required answers included winner, exact score, possession, , confidence, and tactical reasoning.
- 4.7 led overall after 54 matches with 141 points and 9 exact-score picks.
- Sonar 2 had the most correct match winners, with 35.
- For , the practical tip is to keep model settings comparable and judge each prediction category separately.