Most prediction sites publish a confidence number alongside every call. 'HIGH confidence,' '80% lock,' 'AI says ✅.' Almost none of them publish a ledger that lets you check whether those numbers hold up. That's the difference between marketing a number and meaning it.
Calibration in one sentence
A model is calibrated when its stated probabilities match its actual hit rates. If it says 70% confidence on a hundred picks, about seventy of those should win. If it consistently says 70% but hits 55%, the model isn't 'wrong' — it's miscalibrated.
Calibration matters because it's what makes a number actionable. A 70% read on a match is only useful if 70% really means 70%. Otherwise you're trusting a sticker, not a signal.
How we measure ours
Every settled pick on the Cortexscore track record sits in a confidence band — 65–69, 70–79, 80+. The track record page groups settled picks by band and shows the actual hit rate for each. If the 70–79 band trends below 70%, we know the model is over-confident in that range; we'd tighten the publish floor or drop a market entirely.
What HIGH means here
HIGH confidence at Cortexscore means the model surfaces a call only when three independent signals align and there's a clear difference between our read and the market price. The exact rate the HIGH band hits at is on the track record page — including the losses.
Why we publish losses
Hiding losses would invalidate the whole point. The track record exists so anyone — subscriber or not — can audit the confidence labels against reality. Wins, losses, and voids all stay on the page. We don't graduate losing picks out of the ledger.