I've been curious whether you can kill an Claude's text watermark just by editing (not "rewriting/rephrasing") what it wrote. And so I built a Claude/OpenAI/Gemini text-watermark generator plus a detector and threw a bunch of attacks at gpt-oss-20b and Qwen outputs to find out. The technique relies on the famous Tournament Sampling built upon standard Gumbel-max sampling. It turns out almost everything people assume works, doesn't. Swapping em-dashes for hyphens, stripping markdown, ...