A process-grounded framework for submission-time prediction of code review outcomes

Abstract

Relevance. Code review is a central quality assurance practice, yet its outcomes are hard to anticipate at submission: duration varies widely, the number of revision cycles is unknown, and many pull requests quietly lose reviewer attention. Existing approaches address these outcomes in isolation and are predominantly evaluated within a single repository or organization, limiting evidence of cross-project generalization. The purpose of the paper is to establish an explicit model of the code review process and to show that characteristics derived from this model are predictable at submission time by models capable of generalizing across projects. Objectives: to formulate a process model treating automated participants as actors; to derive measurable characteristics of review from its structural elements; to construct a multi-repository dataset; and to validate the predictability of two characteristics under within-project and cross-project evaluation. Methods. The model combines Business Process Model and Notation with a Unified Modeling Language state machine. Gradient boosting classifiers use submission-time attributes of the change, author, and project, computed in an expanding temporal window that prevents data leakage. The empirical study covers 64,820 merged pull requests from 12 GitHub repositories representing six primary programming languages. The evaluation uses stratified cross-validation and leave-one-repository-out validation. Scientific novelty. Prediction targets follow from structural elements of the process model – a state sojourn, a loop, and a timer – rather than pragmatic choice. We treat continuous integration and a large language model assistant as process actors. Results. Iteration prediction generalizes to repositories unseen in training; staleness prediction transfers less reliably, indicating stronger dependence on project context. Practical significance. The proposed models support submission-time risk assessment, review prioritization, early identification of changes likely to require rework, and escalation of pull requests at risk of becoming stale. The findings suggest that iteration prediction can be deployed as a pre-trained cross-project model, whereas staleness prediction benefits from project-specific calibration. Conclusions. Characteristics derived from an explicit process model are predictable at submission time, and a substantial part of this predictability carries across projects. The framework provides a foundation for data-driven review prioritization and future integration of AI-assisted review workflows while preserving human control over review and merge decisions.