A widely shared claim that Claude-assisted commits made rsync buggier does not survive contact with the data, according to an independent analysis of every rsync release with bug records.
Using a severity-weighted bug rate and an exact permutation test, the author found the two Claude releases sit in the middle of the historical distribution. Version 3.4.2 logged zero bugs, the 0th percentile; version 3.4.3 landed at the 77th percentile, elevated but with eight earlier, human-only releases scoring worse. The permutation test returned a p-value of 46%: about a coin flip that any two random releases would look as bad or worse.
The methodology was the author's own, built with a statistician, while the data-fetching scripts were AI-written and every number is templated directly from the analysis to rule out hallucination. The takeaway is not that AI never regresses code. It is that this specific pile-on selected one tail event and narrativised it.