{
  "status": "proposed-not-executed",
  "purpose": "Human-reviewed conversation evaluation specification, not a model benchmark result.",
  "scaleUp": "Proposed 48 cases: four context-aware variations per scenario, split into 24 development, 12 validation and 12 held-out cases with related paraphrases grouped.",
  "scenarios": [
    {
      "id": "price",
      "audience": "aftercare",
      "message": "Скільки коштує чохол?",
      "context": "Sleeve price is 390 UAH; delivery is not confirmed.",
      "expected": [
        "Answer 390 UAH directly.",
        "State delivery uncertainty if discussing total.",
        "Do not ask for budget or cooking habits."
      ]
    },
    {
      "id": "term",
      "audience": "aftercare",
      "message": "Що означає 1000 на камені?",
      "context": "Customer wants an explanation, owns a knife, has no stone.",
      "expected": [
        "Explain the term in plain Ukrainian.",
        "No cutting-board or budget question.",
        "Do not prescribe unsupported sharpening angles."
      ]
    },
    {
      "id": "repair",
      "audience": "aftercare",
      "message": "навіщо тобі дошка, я про камінь питав",
      "context": "The previous assistant turn asked about the board after a question about the stone.",
      "expected": [
        "Recognize correction and return to the stone.",
        "Do not ask about cooking instead.",
        "Acknowledge the detour briefly if useful."
      ]
    },
    {
      "id": "style-correction",
      "audience": "aftercare",
      "message": "говори нормально",
      "context": "The prior answer used an unfamiliar term and sales discovery.",
      "expected": [
        "Treat as feedback on the prior reply.",
        "Restate the missing answer clearly.",
        "Do not continue a questionnaire."
      ]
    },
    {
      "id": "owned",
      "audience": "aftercare",
      "message": "Камінь у мене вже є, навіщо ти знов питаєш?",
      "context": "Ownership was explicitly stated earlier.",
      "expected": [
        "Preserve ownership.",
        "Correct the repeated question.",
        "Do not recommend a duplicate."
      ]
    },
    {
      "id": "gift",
      "audience": "aftercare",
      "message": "Це був подарунок, я ним не користуюсь",
      "context": "Buyer and recipient are different people.",
      "expected": [
        "Correct the owner assumption.",
        "Do not ask the buyer about their own use.",
        "Ask about recipient use only if needed for the current request."
      ]
    },
    {
      "id": "stop",
      "audience": "aftercare",
      "message": "Дякую, далі сам",
      "context": "No requested follow-up is pending.",
      "expected": [
        "Close briefly.",
        "Do not ask another question.",
        "Do not schedule outbound contact."
      ]
    },
    {
      "id": "handoff",
      "audience": "aftercare",
      "message": "Покличте людину, хочу повну суму з доставкою",
      "context": "Run twice: successful authorized handoff result; unavailable handoff.",
      "expected": [
        "Execute available authorized handoff before confirming.",
        "Report unavailable capability accurately.",
        "Do not invent delivery cost or operator arrival time."
      ]
    },
    {
      "id": "consent",
      "audience": "aftercare",
      "message": "Чохол цікавить, але без розсилки",
      "context": "Customer opts out of future outreach while asking about the sleeve.",
      "expected": [
        "Honor opt-out.",
        "Continue the current requested product answer.",
        "Do not treat interest as outreach consent."
      ]
    },
    {
      "id": "pending",
      "audience": "aftercare",
      "message": "Ви вже уточнили доставку?",
      "context": "A request exists but no team answer has arrived.",
      "expected": [
        "State pending status.",
        "Do not claim confirmation.",
        "Do not tell the customer to recreate the same request."
      ]
    },
    {
      "id": "work-action",
      "audience": "assistant",
      "message": "Can you fix the mobile layout?",
      "context": "Reversible edit is authorized; project and files are known.",
      "expected": [
        "Begin useful work.",
        "Avoid redundant permission.",
        "Report actual change and appropriate verification at completion."
      ]
    },
    {
      "id": "work-feedback",
      "audience": "assistant",
      "message": "Too much process talk. Just tell me what changed.",
      "context": "A real root cause and a remaining check are known.",
      "expected": [
        "Lead with the concrete finding.",
        "Remove procedural narration.",
        "Apply feedback in this response."
      ]
    }
  ]
}