NumericGuard

Rejects matches whose prompts do not contain the same numbers.

The highest-value guard in the set, because numbers are where embeddings are weakest. Convert 100 USD to EUR and Convert 250 USD to EUR are ~99% cosine-similar with every mainstream embedding model: the sentences are near-identical and the model was never trained to treat magnitude as meaning. No threshold separates them. Reading the digits does, exactly and for free.

Numbers are compared as a sorted multiset, so re-ordering a prompt does not trigger a rejection, and thousands separators are stripped so 1,000 and 1000 are the same number. A number present on one side only also counts as a difference — that is what catches Explain OAuth 2.0 against Explain OAuth 2.0 to a 5 year old.

A comma is read by what follows it: exactly three digits means grouping and the comma is dropped, anything else means a decimal point and the comma becomes one. So 1,000 is a thousand and 3,5 is three and a half.

Both halves of that rule are load-bearing, and each covers a false hit the other lets through. Dropping every comma turns 3,5 into 35, so "Convert 3,5 km to miles" gets the answer cached for 35 km. Merely splitting on the comma instead turns 3,5 into the two numbers 3 and 5 — which, compared as an unordered multiset, is indistinguishable from 5,3, so "3,5 kg in pounds" gets the answer for 5,3 kg. Only parsing the comma produces one number that differs from both.