How the record is measured

The Model Room is where the product grades itself. Every served standard signal is settled against the real result and written to a graded record, wins and losses alike.

Four cohorts, never pooled

• The standard record: standard lines on an offer a book attested selling, counting every decided row in both directions.
• The recommended standard record: the subset the product actually recommended. This is the cohort worth judging it on, and it is a smaller and harder number than the first.
• The goblin record: reduced payout lines, kept separate because an easier line hits more often and pooling it would flatter the headline.
• The payout ledger: boosted payout demons, which are never recommended, reported with wins, losses and a payout aware figure.

Intervals and pushes

Every block carries a sample size and a confidence interval rather than a bare percentage. Rows from the same slate are correlated, so the interval is reported as a floor on the uncertainty, not a precise one. Pushes leave both the numerator and the denominator, and voided legs write no row at all, so a void count is impossible and is never printed as zero.

Calibration

A hit rate says how often the product was right. Calibration asks a sharper question: when the model claims 70%, does it happen about 70% of the time? The calibration curve and the calibration score answer that per market. A market can hit often and still be badly calibrated, and the product treats the second as the more serious problem.

Trust tiers

Each block carries a computed tier. Graded means outcomes exist. Recalibrated means a measured correction has been applied to that market's probabilities. Market validated would mean the projections beat closing sharp prices, and nothing in the product reaches it, because the stored closing line value is line movement on pick apps rather than sharp closing value.

What the record does not prove

• It is not proof of profitability. A rate must be read against the price each offer was posted at.
• A new engine version resets its own recalibration, so a recently bumped sport has little same version history and says so.
• A projection improved in a replay is not confirmed by that replay. Confirmation is live slates after the change, and the product labels the difference.