Corrections

Some of what is published here will be wrong. When an error is found in something substantive — a claim, a number, a piece of code — the article is fixed and the change is written down here, with what it used to say. Typos and wording are fixed silently.

1. 18 August 2026

   Calibration: when probabilities lie

   It said

   The article said that monotonic calibration never changes the ranking, so the area under the ROC curve is the same before and after.

   It now says

   That holds for Platt scaling and not for isotonic regression. Isotonic is monotonic but flat in places, and flat means ties: fitted on a held-out slice it collapsed 260 distinct scores to 8, and the area under the curve fell from 0.5804 to 0.5667. The article now separates the two methods.

   How it surfaced

   Found while rewriting the article to run against the published sample data. The original claim had been taken from the standard summary of calibration rather than checked.