But wouldn't you have to know the exact context before this token in order to verify the watermark? I.e. a paragraph wouldn't be sufficient; you would need the system prompt, previous prompts, and even hidden thinking?