longPrompts

standard with SubstitutionGuard bounded, for traffic whose prompts carry retrieved context.

M28 measured every guard against prompt length and found one that moves with it. SubstitutionGuard rejects genuine paraphrases at 0% under 48 characters, 12% between 48 and 95, and 15% from 96 characters up — where it flattens and stays at 15% through 512-, 1024- and 2048-character prompts. The step is not a cliff and it is not a slow decay; it is a level change that has already happened by the time a prompt is one line long, and it happens because the guard counts differing positions without ever asking what share of the prompt one position is.

This chain sets SubstitutionGuard.LONG_PROMPT_MAX_TOKENS, so past a dozen content words the guard abstains rather than rejecting on a single word. Everything else is standard.

What it costs, and what it does not fix. The cost is the substitution catches on long prompts, which the README publishes per split. What it does not fix is the other length effect M28 found, because that one is not about length: EntityGuard and DirectionGuard read the first word of the text they are handed as a sentence opener, so a question with retrieved passages in front of it loses that exemption and its opening capital starts counting as an entity. Measured on prompts wrapped in an identical retrieval envelope, entity goes from 6% to 10% and direction from 0% to 4%. No preset can bound that away — it needs the guards to be told where the question starts, which no version of this API can currently say — so it is documented rather than papered over.

longPrompts, reading every marker from vocabulary.