{
  "status": "review",
  "metrics": [
    {
      "metric": "answer_accepted",
      "control_mean": 0.625,
      "treatment_mean": 0.875,
      "delta": 0.25,
      "standard_error": 0.221601,
      "ci95_low": -0.184339,
      "ci95_high": 0.684339
    },
    {
      "metric": "citation_valid",
      "control_mean": 0.75,
      "treatment_mean": 0.875,
      "delta": 0.125,
      "standard_error": 0.205939,
      "ci95_low": -0.27864,
      "ci95_high": 0.52864
    },
    {
      "metric": "retrieval_precision",
      "control_mean": 0.575,
      "treatment_mean": 0.75125,
      "delta": 0.17625,
      "standard_error": 0.042613,
      "ci95_low": 0.092729,
      "ci95_high": 0.259771
    },
    {
      "metric": "latency_ms",
      "control_mean": 848.75,
      "treatment_mean": 1012.5,
      "delta": 163.75,
      "standard_error": 25.786728,
      "ci95_low": 113.208013,
      "ci95_high": 214.291987
    },
    {
      "metric": "cost_eur",
      "control_mean": 0.023125,
      "treatment_mean": 0.031,
      "delta": 0.007875,
      "standard_error": 0.000895,
      "ci95_low": 0.00612,
      "ci95_high": 0.00963
    }
  ],
  "slice_effects": [
    {
      "metric": "answer_accepted",
      "control_mean": 0.5,
      "treatment_mean": 1.0,
      "delta": 0.5,
      "standard_error": 0.5,
      "ci95_low": -0.48,
      "ci95_high": 1.48,
      "query_type": "becas",
      "n": 4
    },
    {
      "metric": "answer_accepted",
      "control_mean": 1.0,
      "treatment_mean": 1.0,
      "delta": 0.0,
      "standard_error": 0.0,
      "ci95_low": 0.0,
      "ci95_high": 0.0,
      "query_type": "matricula",
      "n": 4
    },
    {
      "metric": "answer_accepted",
      "control_mean": 0.5,
      "treatment_mean": 1.0,
      "delta": 0.5,
      "standard_error": 0.5,
      "ci95_low": -0.48,
      "ci95_high": 1.48,
      "query_type": "normativa",
      "n": 4
    },
    {
      "metric": "answer_accepted",
      "control_mean": 0.5,
      "treatment_mean": 0.5,
      "delta": 0.0,
      "standard_error": 0.707107,
      "ci95_low": -1.385929,
      "ci95_high": 1.385929,
      "query_type": "practicas",
      "n": 4
    }
  ],
  "reasons": [
    "el reranker mejora aceptacion y precision, pero aumenta coste y latencia",
    "la metrica citation_valid no mejora de forma perfecta y debe ser guardrail",
    "hay pocos ejemplos por tipo de consulta"
  ]
}