The Method
How the scoreboard is built and graded — analytical and honest, never hot-take culture. Everything below is a rule we hold ourselves to, not a claim about how good we are.
The rules we hold ourselves to
Picks are locked before the walkouts, and never edited
Every pick carries a timestamp written before the fight. Once locked it is never changed, corrected or deleted — audits record what happened, they do not rewrite it. A pick that could be edited after the result is not a prediction.
A missing number stays missing
Where a stat was never recorded you will see a dash, not an estimate. Several 2026 events were locked without a stated confidence — three of them entirely, two more in part — and rather than compute one now and backfill it, those picks are permanently excluded from the calibration table. It makes our sample smaller and our claim smaller with it.
A result that flatters us is treated as a bug
The market is the ceiling we measure against, so an in-house engine that appears to beat it is investigated before it is celebrated. One earlier engine scored better than the market only because it was reading a rating that already contained the result — that is leakage, and it is checked for on every run.
Small samples are labelled, not laundered
A few dozen picks cannot separate a good engine from a lucky one. Where a lead rests on a small sample we say so on the same screen as the number — including when our own larger backtest contradicts the ranking we are showing you.