Show the work.

Only production receipts received by the server before the scheduled game start qualify. Browser records, late backfills, research experiments and older records without pregame proof are excluded. Repeated forecasts for the same exact market and model version count once, using the earliest valid server receipt. Repeated observations remain visible in the audit history.

Hit rate uses all valid WIN and LOSS results. Calibration and Brier use only supported integer-count props at half-point lines, with integer outcomes, where a push is impossible. Their separate calibration sample size is shown. Pending, push, void and inconsistent outcomes are counted separately. Brier is the average squared difference between a frozen probability and its binary result: 0 is perfect, and a constant 50% forecast scores 0.25 on this calibration sample. Integer lines and unsupported probability semantics are excluded from calibration; dropping pushes alone would bias that comparison.

Forecasts from the same game, different lines or different model versions can be correlated. A high hit rate alone does not prove an edge or positive returns. These are observed ledger results, not an independent test set. Automated settlement uses provider final stats; corrections retain revisions. The cutoff fixes which receipts are included; outcomes can update on refresh.