I Let an Open-Source AI Team Refactor a 10,000-Line Legacy Codebase Overnight — Under a Strict Matrix Contract

This article describes an experiment where an autonomous open-source AI team, governed by a deterministic matrix contract, refactors a 10,000-line legacy Flask application overnight. The codebase had no tests, high cyclomatic complexity, and security issues. The author contrasts this approach with naive LLM-based refactoring, which hallucinated dependencies and broke critical functionality. The matrix contract enforces architectural boundaries and validates outputs, preventing hallucinations and ensuring correctness. The article includes the full record of the run, failures, before/after metrics, and a companion repository for reproduction. It is highly relevant to IT/technology, focusing on software engineering, AI-assisted development, code refactoring, and DevOps practices.