Reproducibility Data and Code for Accuracy–Consistency Trade-Offs in Conservation-Constrained and Progressive-Cap PINNs for Inverse Source Identification

This dataset and code package supports the manuscript “Accuracy–Consistency Trade-Offs in Conservation-Constrained and Progressive-Cap PINNs for Inverse Source Identification.” It contains frozen per-run configurations, manifests, run-level metrics, raw and aggregated outputs, statistical analyses, figures, tables, validation and audit records, and scripts for verification and regeneration. The evidence represents 2,464 completed successful rows, including 560 confirmatory factorial runs, reference-grid convergence, finite-difference–Tikhonov audits, optimization and identifiability studies, loss-weight and Pareto confirmation, source-width mismatch, and a held-out inverse-gradient RAR intervention. Large neural checkpoints and dense field arrays are excluded because the reported manuscript results can be verified from the included run-level evidence and regenerated from the frozen configurations. SHA-256 manifests and checksums are included for integrity verification.