MatchGuard

A veto on a candidate match that already cleared the similarity threshold.

Guards exist because cosine similarity is the wrong tool for the last 5% of the decision. Convert 100 USD to EUR and Convert 250 USD to EUR embed at roughly 0.99 with every mainstream model — no threshold you can set will separate them, and serving the wrong one is a silently incorrect answer rather than a slow one. A guard reads the two prompts as text and looks for concrete evidence that the answers differ: a different number, a different unit, a flipped comparison, a swapped entity.

Guards are intentionally asymmetric in cost. A guard that wrongly rejects costs one API call, the exact call you would have made without a cache. A guard that wrongly accepts returns a wrong answer to a user. So a guard should reject on positive evidence, and GuardVerdict.Accept when it has nothing to say — abstaining is always the right move over guessing.

Implementations must be pure, fast (they run on every candidate) and thread-safe.

Inheritors