PhyK-TAS

PhyK-TAS is an experimental decision-support pipeline for assessing transferability risk in spatiotemporal precipitation forecasters. It fuses physical precipitation-regime descriptors, generic distribution-shift diagnostics, and complementary degradation-inference models to estimate whether a source-trained forecaster should be deployed, adapted, or retrained in a target climate regime.

This repository contains the experiment code and derived result artifacts. It intentionally does not contain the LaTeX manuscript.

For reviewer-facing reproduction details, see REPRODUCIBILITY.md. For a direct mapping from paper-facing claims and figures to scripts and files, see TRACEABILITY.md.

Repository Contents

• scripts/: experiment, aggregation, inference, conformal calibration, and figure-generation scripts.
• results/: derived descriptors, shift metrics, transfer-degradation summaries, decision-layer outputs, bootstrap/significance analyses, model-fusion analyses, monotonicity analyses, and conformal decision summaries.
• figures/: generated figures used to inspect and summarize the experiment.
• docs/: local protocol notes, benchmark design notes, and audit summaries.
• REPRODUCIBILITY.md: quick audit path, environment setup, and rebuild scope.
• TRACEABILITY.md: mapping from paper-facing results to scripts and artifacts.

Large raw station archives, generated .npz forecasting tensors, console logs, Python caches, and manuscript folders are excluded from this repository. Several large prediction-level CSV files are also excluded because they are close to or above GitHub's single-file size limit; the aggregate tables needed to inspect the reported results are included in results/.

Main Experiment Artifacts

The current Information Fusion version is centered on the all-viable 11-region benchmark:

• results/all_viable_min100_full_aggregate_report.md
• results/all_viable_min100_full_sanity_report.md
• results/all_viable_min100_full_kbs_results_all.csv
• results/all_viable_min100_full_pair_summary_all.csv
• results/v2_significance_all_viable_min100_full_report.md
• results/v2_meta_models_all_viable_min100_full_report.md
• results/v2_monotonic_all_viable_min100_full_report.md
• results/v2_conformal_all_viable_min100_full_report.md
• results/v2_negative_controls_all_viable_min100_full_report.md
• results/v2_negative_control_delta_all_viable_min100_full_report.md
• results/v2_decision_costs_all_viable_min100_full_report.md
• results/v2_conformal_utility_all_viable_min100_full_report.md
• results/v2_region_sensitivity_all_viable_min100_full_report.md
• results/v3_operational_policy_comparison_all_viable_min100_full_report.md
• results/v3_conflict_confound_control_all_viable_min100_full_report.md

The principal scripts for rebuilding the final analysis tables are:

• scripts/aggregate_experiment_runs.py
• scripts/build_all_viable_final_artifacts.py
• scripts/analyze_v2_significance.py
• scripts/build_v2_meta_models.py
• scripts/build_v2_monotonic.py
• scripts/build_v2_conformal.py
• scripts/build_v2_negative_controls.py
• scripts/build_v2_negative_control_delta.py
• scripts/build_v2_decision_costs.py
• scripts/build_v2_conformal_utility.py
• scripts/build_v2_region_sensitivity.py
• scripts/build_v3_operational_policy_comparison.py
• scripts/build_v3_conflict_confound_control.py
• scripts/build_publication_figures.py

Quick Validation

After cloning the repository, run:

python scripts/validate_release_artifacts.py

The validator checks that core outputs are present, that manuscript folders are absent, and that no included file is near GitHub's 100 MB single-file limit.

Environment

The scripts were developed with Python 3.13 on Windows. Core dependencies are listed in requirements.txt. GPU acceleration is optional for the neural forecasting scripts, but the final aggregation and analysis scripts run on CPU once their input tables exist.

Data Notes

The benchmark is derived from public daily precipitation archives, harmonized outside this repository. The included results/ tables are derived artifacts for the experiment. Rebuilding the complete forecasting pipeline from raw data requires the external station archives and the local harmonization pipeline.

The repository is therefore a reproducibility package for the analysis and derived results, not a mirror of the raw meteorological archives.

Citation And License

Code is released under the MIT License. Citation metadata are provided in CITATION.cff.

Derived tables, documentation, and figures are provided for scholarly reuse with attribution. Raw station archives remain subject to the licenses and access terms of their original public providers and are not redistributed here.

Manuscript Exclusion

The manuscript folders are deliberately not included:

• manuscript_latex/
• manuscript_latex_eswa/
• manuscript_latex_infofusion/

This repository is intended to hold the experiment code and derived artifacts only.