Every prediction we grade, downloadable

This is a contamination-impossible LLM benchmark. Every event happened after every model's training cutoff. Every prediction carries a timestamp from before first pitch. The models never see a betting line: each one gets the identical stats-only packet and returns probabilities, then the games are played. None of it can leak into training data before the forecast exists, which makes this one of the few time-stamped, leakage-free, real-world forecasting benchmarks anywhere. It grows every day.

And you can check our math. The GitHub repo includes a dependency-free Python script that recomputes the entire leaderboard, record, accuracy, Brier, ROI at the closing price, from the CSVs. Run it and you get the same numbers the board shows. That is the whole idea of this site: graded in public, receipts included.