Supply Graph Diagnostic Release

This folder is the clean runnable package for the final SG-24 graph-signal audit paper. It contains:

• data/raw/: the SG-24 raw dataset used by the paper.
• src/focused_audit/: final experiment implementation.
• experiments/: first-class entry points named by the experiment they run.
• focused_audit_config.json: horizons, lag windows, seeds, split ratios, bootstrap settings, and XGBoost search space.

Historical writing-package, revision, cache, and generated-output folders are intentionally not included.

Setup

Run from this folder:

python -m venv .venv
.\.venv\Scripts\Activate.ps1
pip install -r requirements.txt

Run Everything

Full final-paper reproduction:

python experiments\run_all_final_paper_experiments.py --force

Faster smoke run with the same code paths:

python experiments\run_all_final_paper_experiments.py --force --bootstrap-reps 100 --null-draws 3 --shap-sample 100

Individual Experiments

Each script can be run independently. If it needs upstream model outputs, it will create them when missing.

python experiments\run_data_and_graph_summary_experiment.py --force
python experiments\run_selected_split_benchmark_experiment.py --force
python experiments\run_incremental_utility_experiment.py --bootstrap-reps 10000
python experiments\run_feature_utilization_experiment.py --shap-sample 2000
python experiments\run_split_sensitivity_experiment.py --force
python experiments\run_structural_specificity_experiment.py --force --null-draws 30
python experiments\run_disaggregated_error_experiment.py --force

Experiment Map

Final-paper evidence	Script	Main outputs
SG-24 dataset and projected-graph summary	experiments/run_data_and_graph_summary_experiment.py	results/final_paper/final_paper_dataset_temporal_summary.csv, final_paper_graph_summary.csv
Selected-split forecasting benchmark	experiments/run_selected_split_benchmark_experiment.py	data/focused_audit/paper_outputs/main_results_summary.csv, rq1_summary.csv
Level 1 feature utilization	experiments/run_feature_utilization_experiment.py	results/final_paper/final_paper_xai_shap_seed_aggregated.csv, final_paper_xai_within_date_permutation_aggregated.csv, figures/final_paper/final_paper_usage_diagnostics_combined.pdf
Level 2 incremental utility	experiments/run_incremental_utility_experiment.py	results/final_paper/final_paper_date_block_bootstrap_summary.csv
Split sensitivity	experiments/run_split_sensitivity_experiment.py	results/final_paper/final_paper_split_stability_fold_level_summary.csv
Level 3 structural specificity	experiments/run_structural_specificity_experiment.py	results/final_paper/structural_specificity/null_distribution_summary.csv, figures/final_paper/final_paper_null_specificity_distribution.pdf
Product/scaled/regime diagnostics	experiments/run_disaggregated_error_experiment.py	results/final_paper/disaggregated_error/disaggregated_error_summary.csv, regime_error_summary.csv, threshold_sensitivity.csv

Output Policy

Generated outputs are kept out of source control by .gitignore. The clean source and raw dataset remain stable; reruns write derived artifacts under data/focused_audit/, results/final_paper/, and figures/final_paper/.