crowjudge: the bouncer that keeps your cache honest
A cache that serves a wrong answer is worse than no cache. Meet crowjudge, the second model that re-reads borderline matches and vetoes the ones that don't hold up.
Here's the failure mode nobody markets: a semantic cache that's too eager. It sees two questions that look similar, decides they're the same, and serves a confident, wrong answer. For a consumer toy, annoying. For an enterprise, that's a support ticket, a compliance question, or a customer told the opposite of the truth.
A near-match doesn't get served on vibes. crowjudge re-reads both, then decides.
The division of labor is the trick. crowsight is fast and runs on every lookup to narrow a million entries down to a handful of candidates. crowjudge is slower and more thorough, and it only weighs in on the close calls, the ones where a small mistake would be expensive. You get the speed of the embedder and the caution of a reranker, without paying for the reranker on every query.
This is why we can talk about savings and accuracy in the same breath. The savings come from catching every real rephrasing; the accuracy comes from a second opinion that refuses the matches that only looked right. A cache you can't trust isn't a cache, it's a liability with a fast response time.
The bottom line
Speed sells caches; trust keeps them. crowjudge is the quiet part of Crowkis that makes the loud part, 60-70% savings, safe to actually turn on in production.