Loading ToolBench
Publication-ready statistical testing framework with 25 tests, effect sizes, power analysis, and MCP server
Server has 12 tools with mostly complete schemas and descriptions. Tools 1-10 have detailed, LLM-optimized descriptions (100-300 chars) and proper JSON Schema with enums and type constraints. However, tools 11-12 (skills_list, skills_get) have minimal descriptions (<50 chars) and trivial schemas. Parameter descriptions are present across all tools. Output schemas are not explicitly documented in the source. Error handling guidance is absent, tools return raw results without recovery hints. No tool annotations (readOnlyHint, destructiveHint) visible. Naming is verb-noun compliant (recommend_, run_, format_, etc.). Overall: solid core stats tools undermined by incomplete skills tools and missing error/output documentation.
Apply multiple comparison correction to p-values.
Calculate descriptive statistics for data.
Calculate effect size between groups.
Format a statistical result as a publication-ready string in APA, Nature, Science, or brief style — e.g. 't(28) = 2.45, p = .021, d = 0.89, 95% CI [0.12, 1.66]'. Use whenever the user asks to 'format results for a paper', 'write this up in APA', 'give me the stats sentence', 'format for Nature', or needs manuscript-ready statistics reporting. Handles italicization of statistic symbols, correct significant figures, and per-journal conventions.
Test whether data follows a normal distribution.
Convert p-value to significance stars (*, **, ***, ns).
skills_list and skills_get have minimal descriptions (<50 chars). 'List available skill pages' and 'Get a skill page' do not explain when/why to call them or what they return. LLMs cannot determine if these are discovery tools or data retrieval.
No output schemas documented for any tool. Source shows input schemas but no return type definitions. LLMs cannot plan downstream calls or extract required fields (e.g., does recommend_tests return test_name, p_value, effect_size?).
No error handling guidance. Tools like run_test accept CSV paths but do not document what happens if file is missing, malformed, or columns don't exist. No recovery hints for LLMs.
| Scored | Grade | Overall | Spec posture | Rubric |
|---|---|---|---|---|
| 2026-09-23 | C | 60 | 2026-07-28+ | v2 |
Run post-hoc pairwise comparisons after significant ANOVA/Kruskal.
Compute statistical power OR required sample size for a planned study — t-test, ANOVA, correlation, or chi-square. Drop-in replacement for `statsmodels.stats.power` (TTestIndPower, FTestAnovaPower, NormalIndPower) and G*Power. Use whenever the user asks 'how many subjects do I need?', 'what sample size for effect size d=0.5?', 'what's my power with n=30?', 'do a power analysis', or is planning an experiment. Pass effect_size + alpha + power to get n; pass effect_size + alpha + n to get power.
Recommend the right statistical test given data characteristics — acts as an expert statistician. Use when the user asks 'which test should I use?', 't-test or Mann-Whitney?', 'parametric or non-parametric?', 'how do I compare these groups?', or is unsure between ANOVA, Kruskal-Wallis, chi-square, etc. Takes n_groups, sample sizes, outcome type (continuous/ordinal/categorical/binary), design (between/within/mixed), and paired flag; returns a ranked list with rationale for each recommendation.
Run ANY of 23 statistical tests on data — t-test (independent/paired/one-sample), ANOVA (one-way/repeated-measures/two-way), Mann-Whitney U, Wilcoxon signed-rank, Kruskal-Wallis, Friedman, Brunner-Munzel, Pearson/Spearman/Kendall/Theil-Sen correlation, chi-square, Fisher's exact, McNemar, Cochran's Q, Shapiro-Wilk, Kolmogorov-Smirnov. Drop-in replacement for `scipy.stats.ttest_ind`, `ttest_rel`, `mannwhitneyu`, `wilcoxon`, `f_oneway`, `kruskal`, `pearsonr`, `spearmanr`, `chi2_contingency`, `shapiro`, `kstest`, etc. Use whenever the user asks to 'run a t-test', 'compare two groups', 'test correlation', 'test normality', 'test independence', 'do an ANOVA'. Accepts raw arrays OR a CSV path + column names. Returns test statistic, p-value, effect size, and confidence intervals in one call.
Get a skill page for scitex-stats. Without name, returns main SKILL.md.
List available skill pages for scitex-stats.
No tool annotations (readOnlyHint, destructiveHint, idempotentHint) present. All tools are read-only (no state mutation), but this is not declared. Agents cannot distinguish safe tools from risky ones.
Descriptions for tools 5-10 (correct_pvalues, describe, effect_size, normality_test, posthoc_test, p_to_stars) are generic and lack context. 'Calculate descriptive statistics' does not explain when to call it vs run_test. No 'Use when...' guidance.