{
  "model_name": "CESM-SeasonalML",
  "model_type": "cesm_seasonal_ml",
  "architectures": ["CESMSeasonalML"],
  "architecture_metadata": {
    "neural_variants": ["FeedForwardNN", "SeasonalLSTM"],
    "classical_models": ["NumpyRandomForest", "NumpyGradientBoostedTrees"]
  },
  "framework": "PyTorch and NumPy",
  "domain": "earth-science",
  "task": "NDJ/JFM four-class seasonal precipitation forecast",
  "paper": {
    "title": "Training machine learning models on climate model output yields skillful interpretable seasonal precipitation forecasts",
    "doi": "10.1038/s43247-021-00225-4",
    "article_license": "CC BY 4.0"
  },
  "contracts": {
    "rf_xgb_input": ["B", 103],
    "nn_input": ["B", 416],
    "lstm_ndj_input": ["B", 4, 28],
    "lstm_jfm_input": ["B", 12, 28],
    "target": ["B"],
    "classes": 4
  },
  "implementation": {
    "entry_point": "model/cesm_seasonal_ml.py",
    "train_script": "scripts/train.py",
    "inference_script": "scripts/inference.py",
    "evaluation_script": "scripts/result.py",
    "synthetic_data_script": "scripts/fake_data.py"
  },
  "configuration_sources": ["conf/config.yaml", "model/cesm_seasonal_ml.py", "scripts/fake_data.py", "scripts/train.py", "scripts/inference.py", "scripts/result.py"],
  "known_conflicts": [
    "The paper reports 103 RF/XGB predictors but does not provide a complete machine-readable feature manifest that closes exactly to 103.",
    "The last 80 NN dimensions are uniquely named engineering_interaction_* and are a dimension-preserving engineering assumption, not a paper fact.",
    "The paper does not state the final western-US precipitation grid dimensions after preprocessing."
  ]
}