Goodbye to Manual Code Review: The Vibe Coding Experiment Behind TenCirPauli

Description and use

TenCirPauli shows that a fully AI-driven development process can scale to tens of thousands of lines of code, ensuring correctness and performance through automated tests and benchmarks while minimizing manual code review.

Problem

How to develop a large scientific computing project entirely with AI while maintaining quality and performance.

Outcome

TenCirPauli demonstrates that an AI-driven development model can scale to 80k+ lines of code, maintain correctness and performance through automated testing, and reduce human review burden.

Workflow

• High-capability AI analyzes goal and drafts initial Spec
• Human aligns key interfaces, semantics, and architectural trade-offs with Spec
• Medium-capability AI implements the Spec
• High-capability AI reviews implementation and generates Review Report
• Medium-capability AI fixes identified issues
• Automated tests and benchmarks validate correctness and performance
• Process repeats for new features or changes
• Periodic global AI inspections scan repository for accumulated problems

Evidence

The developer’s role is concentrated on goals, specifications, key architectural decisions, and acceptance criteria.

Source

Even medium-capability AI models can now outperform many mature scientific software implementations across a surprisingly broad range of workloads.

Source