{
  "berkeley_bfcl": [
    {"criterion": "Agentic", "weight": 40},
    {"criterion": "Multi-turn", "weight": 30},
    {"criterion": "Live", "weight": 10},
    {"criterion": "Non-live", "weight": 10},
    {"criterion": "Irrelevance", "weight": 10}
  ],
  "mri001": [
    {"criterion": "Single-turn accuracy", "weight": 40},
    {"criterion": "Cost efficiency", "weight": 25},
    {"criterion": "Relevance / irrelevance", "weight": 20},
    {"criterion": "Latency", "weight": 15}
  ]
}
