How the brackets people submitted on /predict stack up, with the model ranked as one more competitor. Every bracket is scored ONLY against actual results: rounds already decided carry points, rounds not yet decided are pending and lock in round by round as results arrive.
Model competitor: the elo-dc-v1.0 model (50,000 sims) committed its PRE-TOURNAMENT bracket once, before kickoff like every human, scored only against actual results. Its current read (conditioned on 104 results, updated 19 Jul 2026) feeds the "Crowd vs the model" comparison only, not its score.
crowd 20% vs result >99% (-80%)
Rank the top three in each group, then tap your way through the knockouts to crown a champion, and see how you stack up against the model.
Pitted against the model's current projected bracket (elo-dc-v1.0).
Every bracket is scored ONLY against actual results. The model competes as one entrant with the bracket it locked BEFORE kickoff (its pre-tournament pick), exactly like every human, so the live re-projection never changes its score. Tap a row for its per-round breakdown: the pills show how many of each round's decided teams the entry also picked. Rounds not yet decided show as pending and lock in as results arrive. Brackets that recorded only group winners + champion are scored on just those dimensions.
This is the model's PRE-TOURNAMENT bracket, locked before kickoff like every human and ranked alongside the human entries, scored against actual results by the same rule · run forecast__elo-dc-v1.0__seed42__cond0__rerun20260610.
A simple round-weighted scheme: deeper correct calls are worth more. Each entry (the model included) earns points for agreement with actual results, with no upset bonuses or tie-breakers (richer leaderboard scoring is a V2 idea, deliberately out of scope pre-launch).