Global scaling, normalized MAE, saved models and marimo workflow

decode_reimplementation_outputs/README.md

@@ -10,8 +10,12 @@ Re-implement the training flow from the DECODE paper for the local I-BLEND data:
 2. Convert original power readings in W to interval energy in Wh.
 3. Merge occupancy and calendar/schedule features.
 4. Optionally merge weather if the local weather range overlaps the energy range.
-5. Add historical energy features, including the previous three same-working-day-class values at the same time instant.
-6. Normalize with Min-Max scaling.
+5. Add historical energy features, including `prev_same_day_type_energy_1`,
+   `prev_same_day_type_energy_2`, and `prev_same_day_type_energy_3`: the
+   energy values from the previous three days that have the same
+   working/non-working class and the same time instant.
+6. Normalize with one shared Min-Max scaler fitted only on the 70% training
+   partitions pooled across all buildings/meters (no validation/test leakage).
 7. Split chronologically into 70% train, 15% validation, and 15% test.
 8. Train Ridge Regression, Decision Tree, Random Forest, and optionally LSTM.
    The LSTM backend prefers PyTorch when available and falls back to TensorFlow/Keras.
@@ -63,6 +67,13 @@ python3 scripts/decode_reimplementation.py --mode meters --target Boys_main --ep
 - `results_paper_buildings.csv`: model metrics for 7 building-level targets.
 - `results_meters.csv`: model metrics for 9 meter-level targets.
 - `run_config_*.json`: run metadata and paths.
+- `global_scalers_*.joblib`: shared global feature/target scalers.
+- `models/<target>/`: fitted sklearn bundles or PyTorch/Keras checkpoints.
+- `predictions/<target>/*_full.csv`: predictions across train, validation and
+  test rows; sequence models begin after their lookback window.
+
+`mae_normalized` is the primary paper-comparable MAE. `mae_wh` is retained as
+an interpretable secondary metric.
 
 ## Notes