# Independent verification notes

All seven declared outputs matched in the offline reference-harness container. A separately implemented least-squares design and state-clustered sandwich covariance calculation also recovered the primary coefficients and confidence interval, without importing the publisher code. This independent check used the pinned analysis table, rather than fresh CDC query results. The returned coefficient interval is broad; the verdict certifies computation, not causal identification or absence of a policy effect.

Hazard screen: none. The inputs contain public state-year mortality aggregates with suppressed small cells, public policy codes, and documentation. They contain no identifiable health records or secrets and supply no mass-harm capability. RAND workbook and archive are public policy materials.

The Benford flags concern bounded rates, selected state-year counts, and overlapping copies of the same measures. Benford conformity is not a validity requirement for these quantities. These flags alone do not demonstrate fabricated data, and I did not certify the source exports merely by matching their analysis.

The code labels its preregistered decision "refute" when a protective association is not significant. That exact computational output matched. It should not be interpreted as statistical proof that the protective effect is zero. Methods and domain review should assess this distinction, staggered adoption, and suppression-driven selection.