Accuracy and corrections

Two different kinds of error can occur here, and they are handled differently.

Editorial errors

These are mistakes in the written guides, glossary and documentation: a wrong figure, an outdated rule, a formula stated incorrectly.

They are corrected in place. The page is fixed and its updated date changes.

Meaningful corrections carry a note. Where a correction changes what a page was telling you to do — as opposed to fixing a typo or tightening wording — a dated line is added stating what was wrong. Pages are not silently rewritten to look as though they were always right.

Rules change, and so do pages. Sports rules, market structures and sportsbook practices all move. Content covering them is reviewed when the underlying thing changes, and seasonal material before each season. The reviewed date on every page is what tells you when that last happened.

Grading errors

These are different in kind, because grades are produced by a model rather than written.

Resolution failures are logged, not discarded. When the engine cannot match a market's entity to history — a team name it does not recognise, a player it cannot identify — it records the failure and returns UNRATED rather than guessing. Those logs are reviewed, and the fixes widen coverage over time.

Grades are not rewritten after the fact. A verdict is a snapshot of what the model computed with the information it had. Retroactively editing them would make the eventual calibration ledger meaningless, since the whole point is to compare what was predicted against what happened.

Systematic errors go on the record. Where a class of grades is found to be biased, the fix goes into the engine and the issue is described on known limitations — which already lists several, including the de-vig bias on lopsided markets and the absence of any weather or game-script variable.

What we do not do

We do not delete pages that turned out to be wrong. They get corrected. A knowledge base that quietly removes its mistakes is not one whose accuracy can be assessed.

We do not adjust historical claims about performance. No track-record claims are made anywhere in the product until the calibration ledger has enough settled bets to support them, precisely so there is nothing to retroactively adjust.

Reporting something

If you find an error, it is worth telling us — factual corrections are prioritised above new content.

Include the URL and what specifically is wrong. Disagreements about interpretation are welcome too, though they are handled as editorial discussion rather than corrections. See about for how to get in touch.