Reproducibility package for reliability auditing of crack-informed spatial corrosion-field posteriors

Description

This reproducibility package supports the analysis of the probabilistic quality of 61-dimensional steel-weight-loss field posteriors conditioned on crack observations. It contains a frozen controlled-test matrix with 6,400 paired evaluation records from two independently generated stage-specific field sets evaluated under 16 mechanism scenarios, archived compact cINN checkpoints, analysis and validation code, machine-readable posterior-quality results, ten-seed crossed-bootstrap outputs, multiplicity-adjusted inference and threshold-sensitivity results, and provenance documentation. The package supports reproduction of the archived analyses from frozen inputs and traceability of the reported results. Complete generation seeds for the historical training and validation data were not retained; therefore, it does not support exact regeneration of those datasets or exact from-scratch retraining of the original checkpoints. The package does not constitute coordinate-aligned real spatial validation or downstream failure-probability analysis. Third-party public specimen source files are not redistributed; their provenance and processing rules are documented separately.

Steps to reproduce

Download and extract the reproducibility package. Create a clean Python environment and install the dependencies listed in requirements.txt. From the package root, run run_validation.ps1 in Windows PowerShell. Alternatively, run validate_package.py and code/analysis/test_exceedance_functionals.py separately. Use provenance/ARTIFACT_MAP.csv to identify the machine-readable files supporting each reported result. Runnable scripts and their required inputs are listed in provenance/CODE_EXECUTION_SCOPE.csv. Any rerun should use a new output directory to preserve the archived results. The package supports reproduction of the archived analyses from frozen inputs but not exact regeneration of the historical training and validation datasets or exact retraining of the original checkpoints.

Institutions

• Anhui University of Science and Technology

  Anhui, Huainan

Categories

Funders

• National Natural Science Foundation of China

  Beijing, Beijing

  Grant ID: National Natural Science Foundation of China, Young Scientists Fund, Grant No. 52308335

Licence