What the published accuracy figure counts, how a prediction is reviewed, and why the misses stay on the site.
Every reviewed prediction is either right or wrong: the side that was picked won, or it did not. The figure on the accuracy page is the first divided by the total. There is no scoring system on top of it and no partial credit.
A published call is fixed. Once the match is played it gets a review that records the outcome and what actually decided it. Nothing rewrites the original probability, the picked side or the reasoning — the page keeps what was said beforehand, which is the only version worth measuring.
A wrong call keeps its page, keeps its number, and is marked as wrong. It counts against the published figure in exactly the same way a correct one counts for it. A record that quietly loses its failures is not a record.
What is stored is the count of right and wrong calls. The percentage is worked out when the page is rendered, so changing how confidence levels are drawn never rewrites history: the old calls keep the level they were reviewed under, and the totals stay comparable across the whole archive.
The same reviews produce a figure per confidence level and a figure per tournament, so the obvious questions have answers: do the higher levels land more often, and is the record carried by one soft event. Both breakdowns come from the same reviews as the headline, not from a separate count.
Over a handful of matches, a percentage is one result and a rounding. Where a sample is too small to mean anything — a team we have called four times — the record is shown as it is, four calls and how they went, and no percentage is printed at all.
Every published prediction that has been played is reviewed. There is no category of match that is quietly left out of the count, and no "unofficial" call that fails to reach the total. The archive lists them, so the count can be checked against the list.
A hit rate over a large sample says the calls have been better than guessing, or that they have not. It does not promise the next one, and it is not a return: a set of correct calls at short odds and a set of incorrect ones at long odds can carry the same percentage and opposite consequences. What a prediction is and is not covers that distinction.