A wrong web answer often appears in none of the retrieved evidence. Skips yes/no and numeric answers; grounds each comma/and-separated part separately. Fires far more often on wrong answers than on correct ones — measure the rate on your own episodes before relying on it. Action: inject "answer X does not appear in any retrieved evidence; cite or search again".