A goals model, a distribution of scorelines, and a nightly grade against the closing line. No part of it is a secret.
Every probability here comes from two figures: the goals we expect the home side to score and the goals we expect the away side to score. For each team the model fits an attack strength and a defence strength on top of its league's own scoring rate and a fitted home-advantage term, weighting recent matches more heavily and blending expected goals with the goals actually scored where xG exists. The technical name is a Dixon-Coles style Poisson goals model with a market-anchored calibration head.
From those two figures we build the full distribution of scorelines and read every market off it: the cells where the home side scores more give the home win, the cells with three goals or more give over 2.5, the cells where both sides score give both teams to score. That is why our markets never contradict each other.
A goals model knows the football but not the news, so the last step before publication pulls each market partway towards the bookmakers' price with the margin removed. The weight is fitted per market on how much that price has helped in the past, and the whole thing refits twice a day.
Where the two disagree by four points or more on an outcome we hold a price for, we mark it in purple. Treat a large gap with suspicion: the bigger it looks, the more likely it is that the market has team news we have not seen. We retired the old label for these, because it promised something we cannot: a bet that is worth making. What we can show is the number and the grade, on the record page. A probability is not a tip. 18+, and please gamble responsibly.