Knockout II — ground-truth
Stage 2 of the experiment. The group stage is over, so every model now starts from the real FIFA results and the official Round-of-32 draw — identical for all — and predicts only the knockout outcomes. Unlike Stage 1 (each model chained on its own predicted groups), this shared starting point makes the brackets directly comparable, tie by tie.
Group-stage retrospective
How each model's group-stage predictions (Stage 1) held up against the real finished group stage — group winners called and real qualifiers it had advancing. The real outcomes (FIFA) are the same starting point every Stage-2 bracket below is built on.
| Run | Group winners | Qualifiers |
|---|---|---|
| OpenAI · web | 11/12 | 27/32 |
| Gemini · web | 11/12 | 26/32 |
| OpenAI · baseline | 11/12 | 26/32 |
| Gemini · baseline | 11/12 | 26/32 |
| Claude · enriched | 10/12 | 26/32 |
| OpenAI · enriched | 10/12 | 26/32 |
| Gemini · enriched | 10/12 | 26/32 |
| Claude · web | 9/12 | 26/32 |
| Claude · baseline | 9/12 | 26/32 |
Stage-2 scoreboard
Every bracket fills the same official Round-of-32 draw, so ties are compared head-to-head: exact score 5, correct winner 3. Survival adds the Stage-1-style stage points (Champion 32).
| Bracket | Ties scored | Winners | Exact | Tie pts | Survival | Total |
|---|---|---|---|---|---|---|
| OpenAI · enriched | 23 | 21 | 3 | 69 | 152 | 221 |
| Claude · web | 22 | 19 | 6 | 69 | 120 | 189 |
| Claude · baseline | 22 | 19 | 4 | 65 | 120 | 185 |
| Gemini · baseline | 23 | 19 | 4 | 65 | 120 | 185 |
| Gemini · web | 21 | 17 | 3 | 57 | 116 | 173 |
| Claude · enriched | 20 | 16 | 5 | 58 | 114 | 172 |
| OpenAI · baseline | 21 | 17 | 4 | 59 | 112 | 171 |
| Gemini · enriched | 21 | 16 | 3 | 54 | 114 | 168 |
| OpenAI · web | 20 | 17 | 1 | 53 | 108 | 161 |
swipe to follow the bracket →