Experiment output structure

All outputs are written inside EXPERIMENT_DIR. The directory is created automatically if it does not exist. After a completed run the layout looks like this:


results/ibd_franzosa/
├── evaluation_results.csv
└── siso/
    └── target-Study.Group/
        ├── cv_config.json
        ├── experiment_progress.db
        ├── fold_diversity_stats.json
        └── taxonomy-{abbrev}/
            └── transform-{type}/
                ├── transform_metadata.json
                └── project-{type}/
                    ├── config.json
                    ├── processed_data.tsv
                    ├── targets.parquet
                    └── models/
                        └── {ModelName}/
                            ├── mpdr.yaml
                            ├── mpma.yaml
                            ├── learner_config.json
                            ├── inner_fold_performance_summary.parquet
                            ├── outer_test_folds_performance_summary.parquet
                            ├── outer_train_folds_performance_summary.parquet
                            ├── predictions/
                            │   ├── outer_test_predictions.parquet
                            │   ├── outer_train_predictions.parquet
                            │   ├── inner_val_predictions.parquet
                            │   └── inner_train_predictions.parquet
                            └── weights/
                                ├── model_R0_OF0.joblib
                                └── ...

The siso/ directory holds single-input/single-output results. Results are grouped first by target, then by taxonomic abbreviation, transformation, and projection. Each {ModelName}/ directory contains the full record for one evaluated combination.

File	Contents
`evaluation_results.csv`	All evaluated configurations with test performance scores, standard deviations, per-fold MCC values, and whether each configuration met the evaluation thresholds. Sorted by test performance score descending.
`cv_config.json`	Cross-validation fold indices used throughout the experiment, saved once per target.
`experiment_progress.db`	SQLite database tracking which configurations have completed, used to resume interrupted runs.
`fold_diversity_stats.json`	Label distribution statistics for each fold.
`config.json`	Configuration snapshot for this taxonomic, transformation, and projection combination.
`processed_data.tsv`	Feature matrix after taxonomic filtering and transformation.
`mpdr.yaml`	MPDR configuration (taxonomic resolution, transformation type, projection) for this model.
`mpma.yaml`	Full MPMA configuration including the base learner name and hyperparameters.
`inner_fold_performance_summary.parquet`	Per-fold inner validation scores used for model selection.
`outer_test_folds_performance_summary.parquet`	Per-fold outer test scores representing the generalisation estimate.
`outer_test_predictions.parquet`	Outer test fold predictions. Required for post-experiment ensemble search.
`inner_val_predictions.parquet`	Inner validation predictions. Required for post-experiment ensemble search.
`model_R{r}_OF{f}.joblib`	Serialised model fitted on repeat `r`, outer fold `f`. Written when `save_model_weights=True`.