{
  "title": "Lead pipe prediction: synthetic evaluation",
  "disclosure": "Entirely synthetic educational companion, not the original TF-IDF/gradient-boosting prototype, field validation, verified material classifications, or evidence of savings.",
  "metrics": [
    {
      "label": "Held-out properties",
      "value": 200
    },
    {
      "label": "Held-out groups",
      "value": 40
    },
    {
      "label": "Shared groups",
      "value": 0
    },
    {
      "label": "Model Brier score",
      "value": "0.1238"
    }
  ],
  "tables": [
    {
      "heading": "Held-out model comparison (lower loss is better)",
      "headers": [
        "Predictor",
        "Brier score",
        "Log loss",
        "Accuracy at 0.5"
      ],
      "rows": [
        [
          "Logistic model",
          "0.1238",
          "0.3951",
          "0.8300"
        ],
        [
          "Training-prevalence baseline",
          "0.2387",
          "0.6705",
          "0.6350"
        ]
      ]
    },
    {
      "heading": "Confusion counts at illustrative threshold 0.5",
      "headers": [
        "Predictor",
        "True negative",
        "False positive",
        "False negative",
        "True positive"
      ],
      "rows": [
        [
          "Logistic model",
          110,
          17,
          17,
          56
        ],
        [
          "Prevalence baseline",
          127,
          0,
          73,
          0
        ]
      ]
    },
    {
      "heading": "Reliability bins (descriptive, not fitted calibration)",
      "headers": [
        "Probability interval",
        "Records",
        "Mean prediction",
        "Observed positive fraction"
      ],
      "rows": [
        [
          "[0.0, 0.2)",
          66,
          "0.0873",
          "0.0455"
        ],
        [
          "[0.2, 0.4)",
          48,
          "0.2935",
          "0.1458"
        ],
        [
          "[0.4, 0.6)",
          23,
          "0.4977",
          "0.4783"
        ],
        [
          "[0.6, 0.8)",
          35,
          "0.7119",
          "0.7714"
        ],
        [
          "[0.8, 1.0]",
          28,
          "0.9084",
          "0.8929"
        ]
      ]
    },
    {
      "heading": "Error slices (overlapping, descriptive)",
      "headers": [
        "Slice",
        "Records",
        "Brier score",
        "False positive",
        "False negative"
      ],
      "rows": [
        [
          "Age present",
          178,
          "0.1194",
          15,
          15
        ],
        [
          "Age missing",
          22,
          "0.1588",
          2,
          2
        ],
        [
          "No conflicting record",
          168,
          "0.1170",
          13,
          14
        ],
        [
          "Conflicting record",
          32,
          "0.1590",
          4,
          3
        ]
      ]
    },
    {
      "heading": "Five largest probability errors (invented held-out properties)",
      "headers": [
        "Synthetic property",
        "Simulated target",
        "Prediction",
        "Age missing",
        "Record conflict"
      ],
      "rows": [
        [
          "synthetic-group-120-property-001",
          0,
          "0.8983",
          "No",
          "No"
        ],
        [
          "synthetic-group-049-property-002",
          0,
          "0.8978",
          "No",
          "No"
        ],
        [
          "synthetic-group-099-property-003",
          0,
          "0.8853",
          "No",
          "No"
        ],
        [
          "synthetic-group-117-property-000",
          1,
          "0.1313",
          "No",
          "No"
        ],
        [
          "synthetic-group-078-property-004",
          1,
          "0.1372",
          "No",
          "No"
        ]
      ]
    }
  ],
  "notes": [
    "All 800 records, group IDs, evidence and target labels are invented; signal is intentionally learnable.",
    "Fixed seeds: data 2026, split 4173. Whole-group 75/25 split; no shared groups or properties.",
    "Features and training settings were fixed before evaluation. No tuning, threshold choice or calibration uses test labels.",
    "Reliability bins describe this one synthetic holdout; they do not establish calibrated probabilities.",
    "The 0.5 threshold is illustrative, not an approved material classification or inspection rule.",
    "Age and record-conflict slices overlap; small slices have unstable estimates. No confidence intervals are claimed.",
    "Real evaluation needs verified labels and spatial, temporal and cross-utility holdouts; this example has none.",
    "This companion is separate from the original professional OCR effort and any historical cost savings."
  ],
  "seed": 2026,
  "settings": {
    "split_seed": 4173,
    "test_fraction": 0.25,
    "threshold": 0.5,
    "epochs": 900,
    "learning_rate": 0.4,
    "l2_penalty": 0.002,
    "calibration_bins": 5,
    "optimizer": "Full-batch gradient descent"
  },
  "split": {
    "train_properties": 600,
    "test_properties": 200,
    "train_groups": 120,
    "test_groups": 40,
    "shared_groups": 0,
    "shared_properties": 0
  },
  "learned_weights": [
    -1.1638613469617853,
    2.7739912035756786,
    1.7700839139114222,
    -0.2605514407532509,
    0.32528095372108057
  ],
  "training_prevalence": 0.4483333333333333,
  "logistic_model": {
    "n": 200,
    "positive_labels": 73,
    "threshold": 0.5,
    "log_loss": 0.39511840052845654,
    "brier_score": 0.12375291512034797,
    "accuracy": 0.83,
    "confusion": {
      "true_negative": 110,
      "false_positive": 17,
      "false_negative": 17,
      "true_positive": 56
    }
  },
  "training_prevalence_baseline": {
    "n": 200,
    "positive_labels": 73,
    "threshold": 0.5,
    "log_loss": 0.6705148332970401,
    "brier_score": 0.23871944444444446,
    "accuracy": 0.635,
    "confusion": {
      "true_negative": 127,
      "false_positive": 0,
      "false_negative": 73,
      "true_positive": 0
    }
  },
  "reliability_bins": [
    {
      "interval": "[0.0, 0.2)",
      "count": 66,
      "mean_prediction": 0.08734186430693329,
      "observed_fraction": 0.045454545454545456
    },
    {
      "interval": "[0.2, 0.4)",
      "count": 48,
      "mean_prediction": 0.2934780683864117,
      "observed_fraction": 0.14583333333333334
    },
    {
      "interval": "[0.4, 0.6)",
      "count": 23,
      "mean_prediction": 0.4976776023475952,
      "observed_fraction": 0.4782608695652174
    },
    {
      "interval": "[0.6, 0.8)",
      "count": 35,
      "mean_prediction": 0.7118684299124317,
      "observed_fraction": 0.7714285714285715
    },
    {
      "interval": "[0.8, 1.0]",
      "count": 28,
      "mean_prediction": 0.9084274883767868,
      "observed_fraction": 0.8928571428571429
    }
  ],
  "error_slices": [
    {
      "slice": "Age present",
      "metrics": {
        "n": 178,
        "positive_labels": 64,
        "threshold": 0.5,
        "log_loss": 0.3819383630596792,
        "brier_score": 0.11942469900760805,
        "accuracy": 0.8314606741573034,
        "confusion": {
          "true_negative": 99,
          "false_positive": 15,
          "false_negative": 15,
          "true_positive": 49
        }
      }
    },
    {
      "slice": "Age missing",
      "metrics": {
        "n": 22,
        "positive_labels": 9,
        "threshold": 0.5,
        "log_loss": 0.5017568855031094,
        "brier_score": 0.15877211821433457,
        "accuracy": 0.8181818181818182,
        "confusion": {
          "true_negative": 11,
          "false_positive": 2,
          "false_negative": 2,
          "true_positive": 7
        }
      }
    },
    {
      "slice": "No conflicting record",
      "metrics": {
        "n": 168,
        "positive_labels": 62,
        "threshold": 0.5,
        "log_loss": 0.3784695502280945,
        "brier_score": 0.11703884072778709,
        "accuracy": 0.8392857142857143,
        "confusion": {
          "true_negative": 93,
          "false_positive": 13,
          "false_negative": 14,
          "true_positive": 48
        }
      }
    },
    {
      "slice": "Conflicting record",
      "metrics": {
        "n": 32,
        "positive_labels": 11,
        "threshold": 0.5,
        "log_loss": 0.4825248646053572,
        "brier_score": 0.15900180568129257,
        "accuracy": 0.78125,
        "confusion": {
          "true_negative": 17,
          "false_positive": 4,
          "false_negative": 3,
          "true_positive": 8
        }
      }
    }
  ],
  "illustrative_uncertainty_review": [
    {
      "property_id": "synthetic-group-006-property-002",
      "synthetic_probability": 0.5012977819626833
    },
    {
      "property_id": "synthetic-group-052-property-003",
      "synthetic_probability": 0.4922135614909196
    },
    {
      "property_id": "synthetic-group-119-property-003",
      "synthetic_probability": 0.510542859327147
    },
    {
      "property_id": "synthetic-group-114-property-003",
      "synthetic_probability": 0.4829728812623101
    },
    {
      "property_id": "synthetic-group-119-property-000",
      "synthetic_probability": 0.4828112862424084
    }
  ]
}