transparent methodology · probabilities, never advice

The Machines

Every match page carries a panel of model voices, each forecasting the same ninety minutes from a different angle. They are real players too: each model converts its own forecast into a pick an hour before kickoff — through the same locked pick system as everyone else — and lives with the consequences on the leaderboard. No hindsight, no edits, no special treatment.

The Statistician

Elo · Poisson · zero vibes

Pure arithmetic. Each team carries a live Elo rating (updated after every final whistle, K = 40, goal-difference weighted; hosts get +60 at home). The rating gap becomes a win expectancy, the win expectancy becomes a goal split around a 2.6-goal match, and two Poisson processes (draws inflated ×1.08) produce the full score grid. The brief it writes is a readout of its own numbers — nothing more.

Joker policy: spends them on its highest-confidence calls. The only bot allowed to.

The Analyst

language model · grounded only

A language model that sees exactly one thing: a context block of verifiable facts — current standings, qualification picture, live ratings, the forecast engine’s probabilities, recent final scores. It is forbidden to invent injuries, quotes or news, and must return probabilities that sum to one (we renormalize server-side regardless). When its gateway is offline, the panel simply runs without it — honestly.

The Crowd

real picks · wisdom & romance

Not a model at all — it’s you. The Crowd’s forecast is the actual distribution of fan picks on each match: outcome shares plus the most-picked scoreline. It stays silent until at least 25 picks exist (and your own view of the distribution stays hidden until you’ve picked — no anchoring). Crowds read form remarkably well. They also, occasionally, pick with their hearts. That’s the experiment.

The disagreement meter

When the voices differ, that’s signal. Each match panel shows a disagreement score — the average distance between the models’ probability splits. Three lit sparks mean the models genuinely can’t agree (distance above 0.25): historically the most interesting matches to watch, and the hardest to call.

What the machines never do

No betting odds, no staking advice, no real-money anything — Bet & Joy is a free-to-play fan platform and the machines output model confidence, full stop. They also never see private data: ratings, results and aggregate pick counts only. The tournament-level engine behind their bigger claims is documented on the forecast page.