Every prediction site publishes a record. Almost none can prove the numbers were not adjusted afterwards — and the tooling makes it trivially easy to do so by accident, never mind on purpose.
Our own site had this flaw for weeks, honestly built. Ratings are recomputed on every data refresh. Improve the model on Tuesday and every historical fixture is re-rated with Tuesday's formula, so a game published at 62% and graded a win might display 58% by Wednesday. Nobody edited anything; the number simply moved. A visitor comparing notes across days would be right to call it revisionism.
Since 30 July 2026 the last data sync before a match kicks off locks that fixture's rating and pick — all four markets, including the cases where we make no pick — into a ledger committed to version control. Graded rows display the frozen numbers permanently. Model improvements apply from the moment of the lock forward and cannot reach backwards. The file is in git, so any change to a past entry is a visible, dated, attributable edit rather than a silent recomputation.
The point is not that we are trustworthy. The point is that the record is checkable without trusting us. A daily commit history of pre-match predictions is a public, timestamped claim; a site that merely displays a strike rate is asking for faith. The same mechanism locks our daily combination selections each morning.
Freezing means our published record cannot benefit from improvements to the model. Ratings that we now believe were slightly wrong stay on the page, graded as they were. That is the correct trade: a record that can be improved retroactively is not a record, it is a marketing surface.
Method: walk-forward testing — every prediction is made using only information available before kick-off, then graded against what happened. Studies run June-July 2026 on the site's own dataset. Questions: contact us.