Module A: Watermark Lab
Generate text carrying a hidden statistical watermark, then test any text for it. Only someone holding the same key can find the signal.
How it works and its limits
Original, independent implementations of two published watermarking algorithms: the Kirchenbauer et al. (2023) green-list scheme, and the Aaronson/Kuditipudi et al. Gumbel (exponential-minimum-sampling) scheme.
Text comes from a small, hand-built toy grammar, not a real language model. The point is the watermark signal, not fluent prose. This cannot detect any vendor's real production watermark. Read the limitations.
Generate
The key seeds which words count as green. Higher bias strength pushes harder toward green words: easier to detect, more distorted text.
Detect
Paste any text. Detection needs the same key, and for green-list the same γ, that generated it.