Statistical Claim Sanity Checker

Sanity-check the statistical claims below before I act on them or forward them. You are not verifying arithmetic against external data — you are auditing the REASONING for the classic ways numbers lie to smart people. Run every claim through this checklist and report only the checks that raise a flag: 1. **Percent of WHAT?** — "43% increase": from what base? A jump from 7 to 10 users is 43%. Is the base disclosed, and is it big enough to mean anything? 2. **Base-rate blindness** — impressive rates on rare events ("doubles your risk" of something affecting 1 in 50,000). 3. **Denominator switching** — does the metric quietly change population between comparisons (revenue per user where "user" was redefined; survey of customers vs. survey of respondents)? 4. **Survivorship** — is the sample only the winners (retention stats excluding churned accounts, "companies that did X grew" ignoring companies that did X and died)? 5. **Seasonality and window-picking** — does the comparison window flatter the trend (March vs. February for anything sold seasonally; "since 2020" baselines)? What would the claim look like year-over-year, or with the window shifted one period? 6. **Correlation dressed as cause** — does the language ("drove", "led to", "thanks to") outrun the design? What confound would produce the same numbers with no causal link? 7. **Aggregation hiding (Simpson's)** — could the overall trend reverse within the obvious subgroups? 8. **Precision theater** — "37.2% improvement" from an n of 30; error bars nowhere in sight. Does the stated precision survive the sample size? 9. **The missing comparison** — a number with no baseline, control, or benchmark ("we resolved 10,000 tickets!"... out of how many, in what time, vs. last quarter?). **For each flag:** quote the claim, name the sin, state what additional number would settle it, and — where possible — the alternative innocent explanation the author should have ruled out. **Close with:** the single most load-bearing claim in the document, and whether it survives. Claims / report: {{claims}}