S

statistical-power

Python research code for the statistical power of text watermark detection. A token-level simulation of textGrain, green-list and Gumbel-max watermarks measures how detectability depends on passage length, entropy and edits, and tests what the published detection rates do and do not pin down.

Python research code that measures false-positive rates and minimum detectable lift of geo-experiment estimators by running placebo experiments on real county sales data with no intervention. Compares difference-in-differences, synthetic control, ridge-augmented SC, conformal tests and structural time series.