What problem does it solve? Results drafts often contain numbers that reproduce but conclusions that do not hold up — causal claims undermined by confounds, striking rates built on tiny subgroups, correlations driven by single outliers. This Skill red-teams every discrete claim in a draft against the underlying dataset before publication, so fragile or wrong claims are hedged, scoped, or retracted instead of shipped. ## Core Features & Use Cases - Claim extraction and reproduction: Splits a prose draft into discrete checkable claims and recomputes each stated statistic directly from the source dataset with Python check snippets. - Adversarial stress-testing: Attacks each reproduced claim with the threats most likely to kill it — outlier sensitivity, Simpson's reversals, confound stratification, subgroup size, and alternative specifications. - Verdict-driven revision: Classifies each claim as verified, fragile, or refuted, then rewrites the draft so fragile claims are hedged or scoped and refuted claims are corrected, with a ledger recording every verdict. - Use Case: Before submitting an A/B test report claiming a treatment improves recovery by 16 points, run this loop to discover the effect reverses within age strata, and revise the draft to a descriptive claim with the causal language retracted. ## Quick Start Verify every claim in results_draft.md against trial_data.csv and produce a revised draft with each claim marked verified, hedged, or retracted.
