> The net change in accuracy does not show whether the same individual calls succeed with and without the watermark.

Why would that matter in the slightest? LLMs aren't deterministic, the same happens with a different seed, or whitespace in the prompt.