any watermarking ai researchers who can explain this?

what if the llm should

- repeat something verbatim (important in a compaction prompt)

- there is just one correct order of tokens for a somewhat long chain (a certain sequence of control signals)

- provide a diff of 2 inputs

without punctuation or whitespace wiggle room?

how does the drifting work?

does it postpone the drifting and drift stronger later?

what if max_tokens is set to a low number?

in what way does this not affect output quality?