These are the figures that held before kick-off: each side's record over every match it had played before 4 Sep 2026, computed exactly as the predictor computes today's. Neither side has an earlier match in the database, so there is nothing to show. Draw score and rank are left out: both place a team against the whole field, and the field can only be scored as it stands now.
The mix is the one that scored best over 544,576 matches — the two are averaged on the log-odds scale, not simply added.
A regression over five signals — the goals model, the draw rate at this Elo gap, the expected goals in the fixture and the competition's own base draw rate — with an isotonic map on top to put the level back where reality is. Held out on a decade it never saw it beat the blend above, which is why both are shown here while one is chosen. Unlike the two readings above it, this one is not leak-free: the model was fitted on all of history, this match included, so for a played match it is a description rather than a prediction.
Both figures are what the models knew before this match — neither ever reads the result it is describing, so they can be compared with what actually happened. Measured over 160,869 priced matches the goals model is slightly worse than the bookmakers (how it is scored), so a high draw chance means an even, low-scoring match rather than value. Descriptive statistics, not betting advice.