ScopeGuard

Rejects matches that ask for the same subject in a different shape.

"Write a haiku about the ocean" and "Write a sonnet about the ocean" differ in one word and share a topic, so they sit close together in embedding space. They are not the same request. Nor are "Give me a code example of a Python decorator" and "Explain the theory behind Python decorators".

What changed in each case is not the subject but the answer being asked for: its format, its length, its depth. Those words are a small, closed set, so the guard reads them directly.

Like EntityGuard and UnitGuard, it rejects a substitution and not an addition — both prompts must name a shape, and the shapes must differ. Firing on one-sided evidence sounds stricter and is simply wrong: "how do I rotate an SSH host key" and "what are the steps to rotate an SSH host key" are one question, and refusing them buys nothing but an extra API call. A genuine depth request that only one side spells out — "how does HTTPS work" against "how does HTTPS work at the packet level" — is the residue this guard does not cover, and is what a dev.kmemo.Verifier is for.