I dont rely only on pinning the judge version. NiriZan keeps a small immutable, human labeled Anchor Set and continuously rescores it alongside production evaluation. The Attribution Engine compares changes in judge behavior on anchors with changes in system behaviour on production metrics
