Our AI picks come from the LLM League — a research experiment in which several large language models predict the same matches under identical rules.
How a model pick happens:
- Each match gets two checkpoints: an early read locked as soon as the match enters our prediction pool (up to ~48 hours out), and a final call locked once the real pre-match information lands — the confirmed lineup and/or fresh team news — or shortly before kick-off if neither arrives in time. At both checkpoints, every active model receives the same prompt with freshly re-fetched match data: head-to-head, recent form, the bookmaker's current 1X2 odds, and any sourced team news collected for that checkpoint. The prompt is versioned and published on the league's methodology block.
- The model must answer in a strict format: an outcome (home/draw/away), its own probability estimate, and a short reasoning note.
- The pick is stored immediately with a timestamp, the frozen odds and a hash of the inputs. It can never be edited afterwards.
- After the final whistle the pick is graded automatically against the 90-minute result.
Models stake flat virtual coins from an equal starting bank, so the league table is a running, public comparison — including each model's misses and skipped matches.
Knockout matches get a second pick. Alongside the 90-minute 1X2 call, every model also answers a separate Winner market for knockout-stage ties: who actually advances after extra time and penalties, no draw option. Both ride on the same checkpoint — no extra prompt, no extra cost — and they can legitimately disagree (a model can call a 90-minute draw while still backing one side on penalties). The Winner pick is shown for transparency and graded publicly, but it's kept out of the bankroll/leaderboard entirely — a deliberately separate, second experiment, not a second shot at the first one.