{
  "title": "Annotvia AI — Arabic evaluation demonstration",
  "sample_type": "synthetic",
  "notice": "Illustrative example only. Not client data, a production export or measured project performance. Schemas and rubrics are agreed per project.",
  "language": "ar-SA",
  "task": "response_preference_evaluation",
  "prompt": "اكتب رسالة قصيرة باللهجة السعودية أطلب فيها تأجيل موعدي لبكرة، بدون ما تحدد ساعة.",
  "responses": {
    "A": "أهلًا، أبغى أأجل موعدي لبكرة إذا ممكن. شكرًا لكم.",
    "B": "أودّ تأجيل موعدي إلى الغد في الساعة الثالثة مساءً."
  },
  "rubric": [
    {
      "criterion": "Instruction following",
      "check": "Requests tomorrow, stays brief, and does not specify a time."
    },
    {
      "criterion": "Dialect and register",
      "check": "Uses natural Saudi Arabic appropriate to a short appointment request."
    }
  ],
  "preferred_response": "A",
  "rationale": "A follows the request, stays brief and uses a natural Saudi expression. B introduces a time the user explicitly excluded and uses a more formal register.",
  "error_annotations": [
    {
      "response": "B",
      "category": "instruction_following",
      "evidence": "في الساعة الثالثة مساءً",
      "explanation": "Specifies a time despite the explicit instruction not to."
    },
    {
      "response": "B",
      "category": "requested_register",
      "evidence": "أودّ تأجيل موعدي إلى الغد",
      "explanation": "Uses formal Arabic rather than the requested Saudi dialect."
    }
  ]
}