prose

standard for traffic that is declarative prose rather than questions.

Register is a property of a deployment, not of a benchmark. A support assistant sees questions, a command palette sees imperatives, a documentation search sees noun phrases, and a retrieval-augmented pipeline compares retrieved passages, which are prose. The chain was tuned on the first of those, and M35 scored every guard per register to find out what that cost.

One guard is responsible for almost all of it. On declarative prose SubstitutionGuard rejects 493 genuine paraphrases to catch 42 near misses, a precision of 8%, where on the written question corpora it runs at 98 to 100% and is the strongest guard in the chain. That is not a tuning problem, it is a mechanism that does not hold on the register: the guard rejects when two prompts have the same content words in the same order and differ in exactly one position, which in a question means a swapped term and in prose means a synonym. So this chain is standard without it, and nothing else changes, because a preset that moved several things at once could not be read.

The cost is real and it is published rather than implied. On the written corpora, which are questions, this gives up protection and buys nothing, which is exactly why it is a preset and not a new default. RegisterPresetTest holds the trade and the README carries the numbers.

prose, reading every marker from vocabulary.