You Spin Me Round: LLM-Assisted Architectural Conformance through Code and Diagram Round Trips

ArchLoop: supplemental material Start here. The package has four folders: Folder Contents tool/ One copy of ArchLoop, installation requirements and selected offline tests data/ Study observations and paper results as CSV files expert-annotations/ The 120-suggestion Excel workbook, individual scores and agreement metrics instructions/ Running instructions, the data index and a result-reproduction script To read the results: open data/paper-results.csv in Excel. Filter table_label to choose a paper table. Each row contains one field from one table row. Labels match the LaTeX source and remain stable if table numbers change. instructions/file-index.csv explains every included file. To inspect expert ratings: open expert-annotations/Expert-annotation.xlsx. Start with Summary, then Expert A and Expert B. Trial ID links each rating to the Suggestions sheet, containing the request, source excerpts and proposed edits. The four criteria contain 960 scores across the two reviewers. To reproduce the saved-data results: with Python 3.11 or later, run: python3 instructions/reproduce.py --output-dir results This needs no additional Python packages, network connection or API key. It checks file hashes, recomputes the study summaries and expert agreement, and checks them against the supplied results. The optional --from-workbook flag also reads the Excel score cells directly; its one dependency is listed in instructions/requirements-analysis.txt. To run ArchLoop: follow instructions/GUIDE.md. This is separate from reproducing the saved results. New model-assisted requests use an OpenRouter key entered locally. No key is included.