Research questionHow can watermarking provide trustworthy provenance for LLM text at scale despite transformations and accumulating false positives?LLM outputs can be copied, transformed, and redistributed without dependable evidence of origin. Watermark signals may weaken through repeated transformations, while even small false-positive rates become consequential when detection operates at large scale.