Four reports filed
Licklider, Inc., the Tokyo company behind SciGround, has detailed four reports that contributed to improvements in research software, the company said. It said the work supports mathematical verification for AI-assisted research, and it credited the maintainers and contributors who implemented and reviewed the corrections.
The R corrections
On 20 perfectly concordant pairs, R's exact Kendall test returned about -4.44e-16 instead of 8.22e-19, and the standard summary printed "p-value < 2.2e-16", concealing the negative number handed to downstream code; taking its logarithm produces NaN. R Core confirmed the fix in development revision c90891 under report #19319. R 4.6.2 remains affected by the Kendall issue although it carries the separate Wilcoxon correction, and the release credits founder Tasuku Kobayashi's reports plus Ahmed Sattar Jabbar's suggestion for that correction. Users of R 4.6.0 and 4.6.1 should update to 4.6.2. Licklider said the examples concern numerical accuracy, not reversals of a 5% significance decision.
Julia and python-flint
HypothesisTests.jl v0.11.8 returned approximately 1.25 for a signed-rank test on [1, 2, -3]; enumerating all eight sign assignments gave 1.0 as the correct result, and the project confirmed a resolution through a separately authored correction released in v0.12.0. In python-flint, a particular constructor input could discard uncertainty for an ordinary decimal such as 0.1, producing an interval that excluded the intended value; the project merged repair #449, and the public report documents affected versions and a workaround pending release.
In the founder's words
"A calculation can finish successfully and still return an invalid result," Kobayashi said. "We independently establish the answer and take reproducible evidence to maintainers, helping improve software other researchers use."