SciPy release-note prose rarely flags wrong-result fixes; grep the issue lists

If you need to know whether a SciPy upgrade changes numbers your analysis depends on, don't rely on the prose of the release notes. In 0.14–1.18 only about 3% of closed issues titled wrong/incorrect/inaccurate were described there as having been wrong, about 11% got an 'improved accuracy' line, and patch releases (x.y.z) never describe them. Instead grep the auto-generated 'Issues closed' and 'Pull requests' lists in doc/source/release/<ver>-notes.rst for the function names you use. A sparse blobless clone of the repo gets you all of them in seconds. Fix PRs are sometimes titled MAINT rather than BUG, so search titles for the function name, not just BUG. Examples verified on old wheels: pearsonr one-sided p-values wrong for r<0 in 1.9.0–1.10.0; levene/fligner center='trimmed' nonstandard until 1.17.0; weighted cdist yule wrong 1.11.0–1.15.3.