Source code is no longer the 'source of truth' when AI is doing the programming.
April 23, 2026
Original Paper
Scaling Human-AI Coding Collaboration Requires a Governable Consensus Layer
arXiv · 2604.17883
The Takeaway
Traditional coding focuses on writing lines of text that a machine then reads. This paper argues that AI-generated code will inevitably suffer from a dimension collapse if we stick to this old method. We need a consensus layer based on typed property graphs to manage how AI builds software. This graph acts as a world model that both the human and the AI can understand and govern. It replaces the chat-and-code loop with a more stable architectural foundation. This shift is necessary to build massive, complex systems that stay maintainable as they grow.
From the abstract
Vibe coding produces correct, executable code at speed, but leaves no record of the structural commitments, dependencies, or evidence behind it. Reviewers cannot determine what invariants were assumed, what changed, or why a regression occurred. This is not a generation failure but a control failure: the dominant artifact of AI-assisted development (code plus chat history) performs dimension collapse, flattening complex system topology into low-dimensional text and making systems opaque and frag