EntityGuard

Rejects matches that talk about different named things.

"What is the capital of Australia" and "what is the capital of Austria" differ by two letters and one continent. Embedding models place them extremely close — both are capital-city questions about a European-sounding place — which makes this the classic false hit. Capitalization is the cheapest available signal that a token names something specific, so the guard compares the capitalized words of the two prompts.

Three rules keep it from misfiring:

The opening capital is ignored. What in "What is Docker" is grammar, not an entity. Only the first word is exempt — see Text.entityTokens for why treating every . : ; as a sentence boundary silently disables this guard on exactly the prompts it exists for.

It rejects on substitution, not on addition. Each prompt must name something the other does not: Austria here, Australia there. One prompt merely mentioning an extra entity is not evidence of a different question — "What does SOLID stand for in OOP?" and "what does each letter of SOLID mean in object oriented design?" want the same answer, and only one of them abbreviates.

Pronouns and initials do not count. English capitalizes I, which would otherwise make every first-person question look like it named something.