{
  "models": [
    {
      "key": "jev",
      "name": "Jev",
      "color": "#a99bff",
      "note": "Jev needed little adjustment on this dataset.",
      "raw": [
        {
          "predicted": 0.23200000000000004,
          "observed": 0.3,
          "n": 20
        },
        {
          "predicted": 0.5845,
          "observed": 0.45,
          "n": 20
        },
        {
          "predicted": 0.7809999999999999,
          "observed": 0.75,
          "n": 20
        },
        {
          "predicted": 0.8789473684210526,
          "observed": 0.7894736842105263,
          "n": 19
        },
        {
          "predicted": 0.9357894736842104,
          "observed": 0.8421052631578947,
          "n": 19
        }
      ],
      "adjusted": [
        {
          "predicted": 0.24764093261228295,
          "observed": 0.3,
          "n": 20
        },
        {
          "predicted": 0.5240973312751357,
          "observed": 0.45,
          "n": 20
        },
        {
          "predicted": 0.6955336373466834,
          "observed": 0.75,
          "n": 20
        },
        {
          "predicted": 0.7963044669890706,
          "observed": 0.8947368421052632,
          "n": 19
        },
        {
          "predicted": 0.8704796315097422,
          "observed": 0.7368421052631579,
          "n": 19
        }
      ]
    },
    {
      "key": "perplexity",
      "name": "Perplexity",
      "color": "#6cc9d9",
      "note": "Perplexity underestimated failure risk in several groups.",
      "raw": [
        {
          "predicted": 0.014874422551857206,
          "observed": 0.3,
          "n": 20
        },
        {
          "predicted": 0.10872569989905231,
          "observed": 0.4,
          "n": 20
        },
        {
          "predicted": 0.4315332717161577,
          "observed": 0.7,
          "n": 20
        },
        {
          "predicted": 0.7625982431847347,
          "observed": 0.8947368421052632,
          "n": 19
        },
        {
          "predicted": 0.9299391749435667,
          "observed": 0.8421052631578947,
          "n": 19
        }
      ],
      "adjusted": [
        {
          "predicted": 0.24316729973969622,
          "observed": 0.3,
          "n": 20
        },
        {
          "predicted": 0.4865554315149054,
          "observed": 0.4,
          "n": 20
        },
        {
          "predicted": 0.6893317658830009,
          "observed": 0.75,
          "n": 20
        },
        {
          "predicted": 0.8183657912652641,
          "observed": 0.8947368421052632,
          "n": 19
        },
        {
          "predicted": 0.8984534098823759,
          "observed": 0.7894736842105263,
          "n": 19
        }
      ]
    },
    {
      "key": "liquid",
      "name": "Liquid d1",
      "color": "#edc66b",
      "note": "Liquid d1 underestimated failure risk in several groups.",
      "raw": [
        {
          "predicted": 0.048715954315989074,
          "observed": 0.25,
          "n": 20
        },
        {
          "predicted": 0.16849222817122356,
          "observed": 0.65,
          "n": 20
        },
        {
          "predicted": 0.3353541850302371,
          "observed": 0.7,
          "n": 20
        },
        {
          "predicted": 0.5946524900721527,
          "observed": 0.7368421052631579,
          "n": 19
        },
        {
          "predicted": 0.801753421923037,
          "observed": 0.7894736842105263,
          "n": 19
        }
      ],
      "adjusted": [
        {
          "predicted": 0.32460373721551716,
          "observed": 0.25,
          "n": 20
        },
        {
          "predicted": 0.5429948971146557,
          "observed": 0.65,
          "n": 20
        },
        {
          "predicted": 0.6545894185764363,
          "observed": 0.7,
          "n": 20
        },
        {
          "predicted": 0.7651876518905374,
          "observed": 0.7368421052631579,
          "n": 19
        },
        {
          "predicted": 0.8451709254840251,
          "observed": 0.7894736842105263,
          "n": 19
        }
      ]
    },
    {
      "key": "liquid_ultra",
      "name": "Liquid d1-ultra",
      "color": "#f09b65",
      "note": "Liquid d1-ultra needed little adjustment on this dataset.",
      "raw": [
        {
          "predicted": 0.10561187370434458,
          "observed": 0.25,
          "n": 20
        },
        {
          "predicted": 0.3743228805994703,
          "observed": 0.4,
          "n": 20
        },
        {
          "predicted": 0.6687235779636953,
          "observed": 0.75,
          "n": 20
        },
        {
          "predicted": 0.8346832282501757,
          "observed": 0.8421052631578947,
          "n": 19
        },
        {
          "predicted": 0.9169609266901734,
          "observed": 0.8947368421052632,
          "n": 19
        }
      ],
      "adjusted": [
        {
          "predicted": 0.19924481505206423,
          "observed": 0.25,
          "n": 20
        },
        {
          "predicted": 0.4830318537288673,
          "observed": 0.4,
          "n": 20
        },
        {
          "predicted": 0.7126361520784378,
          "observed": 0.75,
          "n": 20
        },
        {
          "predicted": 0.8354866743350289,
          "observed": 0.8421052631578947,
          "n": 19
        },
        {
          "predicted": 0.9037645007424091,
          "observed": 0.8947368421052632,
          "n": 19
        }
      ]
    },
    {
      "key": "clef_flash",
      "name": "Clef Flash",
      "color": "#e894b1",
      "note": "Clef Flash: one group predicted 8% failure risk, but 60% contained failures.",
      "raw": [
        {
          "predicted": 0.046669999999999996,
          "observed": 0.35,
          "n": 20
        },
        {
          "predicted": 0.07747,
          "observed": 0.6,
          "n": 20
        },
        {
          "predicted": 0.10643499999999999,
          "observed": 0.6,
          "n": 20
        },
        {
          "predicted": 0.19611052631578949,
          "observed": 0.6842105263157895,
          "n": 19
        },
        {
          "predicted": 0.6209894736842105,
          "observed": 0.8947368421052632,
          "n": 19
        }
      ],
      "adjusted": [
        {
          "predicted": 0.38859679233351635,
          "observed": 0.35,
          "n": 20
        },
        {
          "predicted": 0.5073338424955067,
          "observed": 0.6,
          "n": 20
        },
        {
          "predicted": 0.5874823519627255,
          "observed": 0.6,
          "n": 20
        },
        {
          "predicted": 0.7239389375766031,
          "observed": 0.6842105263157895,
          "n": 19
        },
        {
          "predicted": 0.9254919243691905,
          "observed": 0.8947368421052632,
          "n": 19
        }
      ]
    },
    {
      "key": "openai",
      "name": "OpenAI Decisions",
      "color": "#87c99b",
      "note": "The Platt fit shown here did not improve OpenAI’s probabilities.",
      "raw": [
        {
          "predicted": 0.013500000000000002,
          "observed": 0.25,
          "n": 20
        },
        {
          "predicted": 0.47750000000000004,
          "observed": 0.25,
          "n": 20
        },
        {
          "predicted": 0.8699999999999999,
          "observed": 0.75,
          "n": 20
        },
        {
          "predicted": 0.94,
          "observed": 0.9473684210526315,
          "n": 19
        },
        {
          "predicted": 0.9942105263157895,
          "observed": 0.9473684210526315,
          "n": 19
        }
      ],
      "adjusted": [
        {
          "predicted": 0.16299876579412226,
          "observed": 0.25,
          "n": 20
        },
        {
          "predicted": 0.58917727102631,
          "observed": 0.25,
          "n": 20
        },
        {
          "predicted": 0.7083023616058914,
          "observed": 0.85,
          "n": 20
        },
        {
          "predicted": 0.7496144920273857,
          "observed": 0.8421052631578947,
          "n": 19
        },
        {
          "predicted": 0.9152280272411288,
          "observed": 0.9473684210526315,
          "n": 19
        }
      ]
    }
  ]
}