K-Dense-AI skills test: Claude's answer held, support improved

The test was run as day 17 of a daily AI-news verification challenge, following a CLAUDE.md rule to avoid candidates that require paid API keys. That narrowed the 177-skill repository to statistical-analysis (built on Pingouin, SciPy, statsmodels and PyMC) and literature-review (using public PubMed and arXiv APIs). Setup itself surfaced two environment issues: the install command was blocked by Claude Code's auto mode classifier as external code execution until a human allowed it, and installing via Windows Node.js from a WSL UNC path failed with an EPERM mkdir error because the CLI fell back to C:\Windows; running from a non-UNC Windows path succeeded.

The statistical comparison used a fictional pilot with an outlier in Group A. In the no-skill condition Claude Code cited outlier, unequal-variance and small-sample reasons and already named Mann-Whitney U, but gave no test statistic or p-value. With the skill, assumption_checks.py ran and produced descriptives, a Shapiro-Wilk check (Group A p=0.000127, not normal; Group B p=0.248496, normal), a Levene result (p=0.241, variance_ratio=224.37), a Welch t-test and a Mann-Whitney U result, then a sensitivity run excluding the 55. The conclusion matched, but the report now included the figures and a caveat that a non-significant Welch result at n=6 should not be read as no effect.

The literature-review side showed a structure not visible from the README: the bundled script is a post-processing tool, not a search engine. When arXiv returned 0 hits for 'Mann-Whitney AND outlier', Claude Code reworded the query across four patterns and ended with 8 hits under cat:stat.ME AND abs:"Mann-Whitney" AND abs:robust. The skill's own report noted that few papers center on the word outlier and that the hits were adjacent concepts such as robustness and heavy tails.