Replication Package for: "Understanding the Rejection of AI-Generated Code in Pull Requests

Replication Package: Understanding the Rejection of AI-Generated Code in Pull Requests This repository contains the supplementary material, datasets, and analysis scripts necessary to replicate the findings of the paper "Understanding the Rejection of AI-Generated Code in Pull Requests", submitted to the 40th Brazilian Symposium on Software Engineering (SBES 2026). ## Repository Structure and Contents This package includes the following artifacts, aligned with the methodology presented in the paper: * **`SBES_AI_Rejection_2026.ipynb`** The primary Jupyter Notebook containing the data mining and filtering scripts. It details the process of querying the AIDev-pop dataset to identify rejected pull requests (state 'closed' and 'merged_at' is null), filtering out bot interactions, and performing the 10% stratified random sampling. * **`export.html`** An HTML export of the Jupyter Notebook analysis for accessible viewing without requiring a local Python/Jupyter environment setup. * **`human_rejected_comments_with_task_type.csv`** The intermediate dataset containing the 7,270 filtered pull requests that were explicitly rejected. This dataset retains only the final comments authored by human developers (excluding automated bots), mapped alongside the specific task type and the AI agent that generated the PR. * **`amostra_rejeitada_proporcional.csv`** The final subset comprising the 372 pull requests (the 10% stratified random sample). This is the exact dataset used by the authors to conduct the manual qualitative review. * **`Coding.xlsx`** *(or .csv)* The qualitative codebook and thematic analysis mapping. This file contains the raw data analysis, including the extracted quotes from maintainers, the initial descriptive codes assigned by the researchers, and the final aggregation into the four themes comprising the "Hierarchy of Rejection" (Strategic Misalignment, Implementation Gaps, Architectural Incompatibility, and Subjective Repulsion). ## Usage Researchers aiming to replicate the quantitative filtering should refer to the `.ipynb` script and the initial 7,270 PRs dataset. Those interested in the qualitative foundations of the "Hierarchy of Rejection" should examine the `Coding` spreadsheet alongside the 372 sampled PRs.