Tag
The paper analyzes statistical watermarks in language models by formalizing meaning-preserving transformations using linguistic holonomy, proving that watermark detection depends on the survival of seeding windows and establishing a decay law for the watermark signal.