H

hypothesis-testing

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.