Every morning, after the real tips are done, we ask the AI a harder question: rate every runner in every race with the betting odds deliberately removed. No prices, no market moves, no betting forecast — just the form book. Its answers are recorded before the races run and scored here, winners and losers alike.
Why? The AI estimate on our pick cards is honest but anchored: the model sees the odds when it picks, so its percentages track the market closely. This page is the opposite experiment — the model’s genuinely unaided opinion. Our own testing says an AI’s blind read is worse than the market, which is exactly why the site never bets on it. This page checks that verdict in public, live, either way it goes.
First capture day is coming up — predictions will appear here the morning they're made.
No verdict yet, by design: the read-out rules were fixed before the first capture. The strike-rate comparison is quoted at 100 settled races (0 so far); the calibration table below unlocks at 1,000 scored runner predictions (0 so far).
The rules of this experiment were fixed in writing before the first capture: the model, the price-stripping, the calibration bands and the sample sizes above cannot be changed to flatter the result. Predictions are stored with a timestamp before each race’s off time; a race that has already started is never captured. The blind percentage also appears on the day’s pick cards as the “Blind read” cell — shown as the AI’s opinion, never as a value signal — and this page is where it is scored. It plays no part in how the site’s actual tips are chosen, and it is not betting advice. The “Market %” column is the field’s de-vigged probability from the best bookmaker odds at capture time.
