We added Jev to Maison Safqa yesterday. It doesn’t replace Luna, our slower product-page check. Jev decides whether a candidate URL is worth sending to Luna at all: candidate product page → Jev relevance gate → Luna checks only the pages that may hold evidence On a replay of 60 production suggestions / 326 candidate URLs: • 72 Luna page checks avoided • 105/105 evidence URLs kept • no candidate lost its entire evidence set Jev returns a fast, typed probability. Our code still owns the threshold and the workflow. If Jev fails or gives a malformed answer, we fail open and Luna still gets the page. Small decision model. Expensive model only when it earns the call.

Horus had a bug where the error count showed failures but the grouped results were empty. Logs without event_code were dropped from the aggregation. The fix was a missing-value bucket. "No groups" and "no errors" are different states. github.com/Meritt-dev/horus/…