Reproducibility Package for "Validating LLM-Generated Monolith Decompositions with Structural Evidence from Source Code"

This repository contains all artifacts required to reproduce the experimental evaluation described in the paper "Validando Decomposições de Monólitos Geradas por LLMs com Evidências Estruturais do Código-Fonte" (submitted to the Software Engineering course). The study assesses the structural adherence of microservice decompositions produced by the Gemini gemini-pro-latest large language model, using zero-shot and few-shot prompting strategies. Two open-source Java systems—Spring PetClinic and Bookstore—are used as subjects. The package includes: Generation scripts: gemini_agent.py, which calls the Gemini API to produce 30 independent decomposition CSV files for each combination of system and prompting strategy (120 files total). Validation pipeline: Python scripts (violation_detection.py, static analysis modules) that extract dependency graphs from source code using Tree-sitter and NetworkX, map classes to services, detect structural violations, and compute the metrics TVD (Dependency Violation Rate), TPD (Dependency Preservation Rate), and granularity. Batch evaluation scripts: batch_bookstore_normalized.py and batch_petclinic_normalized.py that process all generated CSV files, produce raw metric CSVs (30 runs per strategy), and compute normalized metrics by intersecting the class-to-service mappings of the two strategies to control for coverage bias. Input data: source code of the monoliths (PetClinic and Bookstore), textual requirements, and few-shot examples. Intermediate outputs: dependency graphs in GraphML format, edge lists, and class-to-service mapping tables. Compiled results: raw metrics CSV (resultados_bookstore_30_runs.csv, resultados_petclinic_30_runs.csv) and normalized metrics CSV (resultados_bookstore_30_runs_normalized.csv, resultados_petclinic_30_runs_normalized.csv). Statistical analysis: scripts and a Markdown report (Statistical_Results.md) that present descriptive statistics, Mann‑Whitney U tests, confidence intervals, and effect sizes, along with the pgfplots code for the figures embedded in the paper. Instructions: a README.md file detailing the required environment (Python 3.8+, Tree-sitter, NetworkX, Pandas, google-generativeai) and step-by-step commands to regenerate decompositions, run evaluations, and reproduce the statistical analyses. All artifacts are organized to facilitate reuse and extension with other systems or LLMs.