On-Device GitHub Contributions Audit utilizing Recursive Language Models

Abstract

Senior engineers must verify what developers built, but evidence is buried in commit histories, too voluminous to review manually. This code is often confidential, ruling out cloud services, yet local models cannot fit a repository in context. We present an on-device Recursive Language Model (RLM) for GitHub contribution auditing. A Thinker decomposes the audit into sub-questions and plans traversal, while a smaller Worker reads relevant code slices; the two recurse until evidence suffices. The system checks claims against commit evidence, returning confirmed, unsupported, or contradicted verdicts plus authorship attribution, locally. We evaluate verdict quality, attribution, and agentic quality across four Thinker–Worker pairs, plus a benchmark isolating recursion. Compared to singular model testing result, RLM outperforms them at 0.91 and 0.88 in accuracy.