PaveFEA-AI replication archive for "From generated scripts to solved models: auditing specification conformance in an LLM-assisted finite-element workflow for pavement structures"

Replication archive for the manuscript "From generated scripts to solved models: auditing specification conformance in an LLM-assisted finite-element workflow for pavement structures" (Y. Luo, D. Huo, J. Wang; Chongqing Jiaotong University). It holds the run ledger of 377 logged generations with per-request seeds and request order; all 377 generated Abaqus scripts; the frozen evaluator and scoring rubric; the prompt templates and the 310 rendered prompts; per-generation retrieval metadata; the frozen test-case specifications; the manual repair chains with per-step diffs and the seven repaired input decks; the repair lineage of the seven executed scripts, classifying every repair step by whether it changed the physical definition of the model; the specification-conformance census of the solved decks with its frozen rule set; the synchronisation-timing probe records; the statistical recomputation pipeline and the reproducibility gate; and a snapshot of the submission (manuscript, supplementary material, figures, data tables and verification reports) sealed by RELEASE_SHA256.csv. MANIFEST_SHA256.csv lists the SHA-256 of every file in the archive. What is withheld: the retrieval base derives from a proprietary finite-element function library. In each rendered prompt the retrieved-chunk block, and only that block, is replaced by a manifest of retrieval order, module, function name, similarity score and characters withheld (2,733,580 characters across the 310 prompts). Output databases and solver scratch files (103 files, 1.73 GB) are listed by path, size and SHA-256 in WITHHELD_LARGE_ARTEFACTS.csv and are available from the corresponding author on request. See ARCHIVE_README.md. Version 1.3 (2026-09-29) updates the submission snapshot to the revised manuscript, now titled as above (versions 1.0–1.2 carried earlier states under the title "Static and executable quality come apart in LLM-generated finite-element scripts for pavement structures: an execution-grounded validation"). It adds submission_pack/data/repair_lineage/ (Supplementary Table S17), whose builder runs in this archive's layout as well as in the project repository, and relabels two design descriptions in the statistics output (stats_v1.json) without changing any value. Two internal working documents of the submission package (in Chinese) remain omitted, with their SHA-256 values kept in RELEASE_SHA256.csv. The experimental artefacts are unchanged from versions 1.0–1.2. Archive file: pavefea-ai-replication.zip (2,003 entries), SHA-256 5ea96186e898f33e684eb9b08c919f2570470b262320ca567841256f2edde863.