Reproducibility package: Evaluation design shapes the reported fidelity of urban-flood surrogate models across forecast leads

Reproducibility package for the manuscript "Evaluation design shapes the reported fidelity of urban-flood surrogate models across forecast leads" by Byung Jin Lee and Yoon-Seop Chang (revised manuscript, under review). The package reruns the statistical analysis of the manuscript from evaluation sufficient statistics and checks every table and quoted number against the expected outputs. It does not train models, run the hydraulic model, or expose cell coordinates. Contents- evaluation/: sums and contingency counts written by the evaluators for every model, seed, lead, and scope, including the flood-focused and fixed-hotspot weights, mask-width variants, the self-initialized experiment, and crop-inference results.- data/: persistence sufficient statistics, the held-out rainfall combinations, event-level FVM metrics of the 144 design storms, and seed-specific split assignments.- code/: Python analysis scripts, run_package.py, and verify_package.py.- expected/: expected tables, presentation-level numbers, and hydraulic-response results.- README.md, DATA_DICTIONARY.md, requirements.txt, SHA256SUMS.txt, and MANIFEST_SHA256.txt (checksums of all files). Usage (Python 3.12): pip install -r requirements.txt; python code/run_package.py; python code/verify_package.py. The verification prints "ALL OUTPUTS REPRODUCED" when all 96 compared outputs match. Not included: raw SWMM–FVM inputs and outputs, GIS and drainage-network data, the hydraulic solver, trained model weights, cell-level predictions, and coordinates, which cannot be redistributed because of contractual, institutional information-security, and intellectual-property restrictions.