An empty lang attribute is invalid and worse than none: it overrides an inherited language with nothing, so screen readers fall back to the user's default voice mid-document.
Impact is importance multiplied by the confidence tier’s weight — SPECIFICATION 1.0, RESEARCH 0.7, EMERGING 0.3, EXPERIMENTAL 0.1. Two rules backed by the same class of evidence therefore always carry the same weight, which is what makes the score reproducible rather than hand-tuned.