shortQuestions

standard with SubstitutionGuard reading three-word questions, for traffic that is mostly short.

M51 measured what the substitution floor costs on short prompts, which is most of what real question traffic is made of: on the external question split the median near miss has four content words after stopword removal and a fifth have fewer, and at a floor of four the guard never sees them.

Lowering the floor outright was measured and declined. What this chain sets instead is SubstitutionGuard.withHeadFloor, which applies the old floor only to a difference in the first content word, because the head is where a question keeps its verb and a verb is what the floor was written about.

The trade, published rather than implied. Against standard it costs the tuned split nothing, gains catches on both retired splits for no paraphrase, and on the external question split takes the chain from 65.4% to 68.0% of near misses caught while taking paraphrases kept from 78.9% to 77.6%. Sixty-five wrong answers stopped for thirty-six extra API calls. Whether that is worth it is a question about your domain, which is why it is a preset and standard is untouched. SubstitutionFloorTest holds the numbers.

shortQuestions, reading every marker from vocabulary.