Before every game the scores page shows a win percentage on each side. This is what happened to those numbers: every game, in order, green when the team BEACON had over 50% won and red when it lost. Nothing is edited after the fact, including the misses.
How to read it. Every call is grouped by the confidence BEACON stated before kickoff. If the games it called at 80-89% are won about 85% of the time, the number means what it says. A gap well below zero means the model is over-confident; well above, under-confident. Small samples swing; read the Games column first.
How to read it. The straight-up record cut by whether BEACON called the market favourite or the underdog, the home or the away team, by the model version that made the call, and by the conference of the team it called. Games is the sample; read it before the percentage.
How to read it. In the Win % lens, Record is BEACON's winner call on every game it logged that week; in the Bets lens it is the plays (Conviction and Strong markets). Running is the total from the first graded game through the end of that week.
Bet or not, it is recorded. Every board game, every market, from the last snapshot before kickoff: BEACON's lean on each market (Skip tier included, which was never a play), HUNCH's side where it was scored, the market favourite, and whether the two engines agreed. Units are one per market at the stored price. The miss line is the mean distance of our projected margin and total from the final against the market line's own distance: lower than the book means the model projects the game better than the market prices it. Open Full ledger on any game below for every number it carried.
Not the record. For games BEACON never called before kickoff, this is what the model would have said: the ratings ledger's pre-game probability (walk-forward from results before the game, recomputed with today's parameters) and, where one was logged before kickoff, the market spread from the line history. It is kept apart from the record above and is never added to it. Games with a published call are excluded here so nothing is counted twice.
The call. Before kickoff, a cron logs every upcoming game with the win percentage the scores page is showing at that moment: the rating model on its own, before any market adjustment. The row exists before the game does. After the final, a second job attaches the score and marks the call won, lost, or pushed. The last log before kickoff is the one graded, and the page tells you how many hours before kickoff it was written.
The benchmark. A model that only ever agrees with the betting favourite will match the market's record and prove nothing. So the report card shows the market favourite's record on the same games, and BEACON's record on the games where it took the other side. That second number is where a model earns or loses its credibility. Games the market never priced are counted too, on their own line.
What counts. Every game that had a percentage on the page. A 53% call is a near coin flip; it still counts as a call, and a miss is a miss. Ties are pushes and are excluded from the percentages. The Bets lens shows the same games through the betting markets, for readers who want that; it never changes the win-percent record.
Where the numbers come from. The public record API reads the prediction log's rating-call table and computes nothing from a final score except the grade. Games with a betting line have been logged since 2026-09-18; every upcoming game, line or not, since 2026-09-20.