What a reproducible evaluation requires
Record model version, issue time, target horizon, settlement venue and price, and a probability for a precisely defined outcome before it happens. Freeze the test period and include every eligible forecast. Report sample size, exclusions, wins, losses and ties. Brier score is mean((p − y)²), where p is an outcome probability and y is 0 or 1. A category confidence alone is insufficient. Publish calibration bins and uncertainty alongside scores; do not substitute training fit for out-of-sample performance.