WMTrace

Forensic detection, benchmarking, and explainable analysis of LLM text watermarks, provenance signals, and hidden-payload channels.

WMTrace answers a narrow question — does this text carry calibrated evidence for a declared watermark scheme, key, and configuration? — and states, with every result, what that result does not prove.

A negative watermark result does not mean text was human-written.

Source on GitHub

Writing

• I Tried to Steal a Watermark Key Without Having It — one scheme handed its favored set over at 3.5× chance, the other gave up nothing, and the first explanation I wrote for why turned out to be noise.