Leaderboard snapshots stay blind-friendly but statistically honest.

Ratings are grouped into all-time, 30-day, and 7-day windows. Models with thin sample counts stay provisional until enough votes and sessions accumulate.