Lf-lean: The frontier of verified software engineering

5 hours ago

lf-lean is a verified translation of 1,276 statements from the Logical Foundations textbook from Rocq to Lean, achieved with AI and minimal human effort.
The project demonstrates a 350x speed-up in verification effort compared to manual methods, reducing human oversight from O(n) to O(1).
Task-level specification generators enable correctness to be defined once for an entire class of tasks, allowing automatic verification across instances and codebases.
The approach leverages semantics-preserving transformations, making it scalable regardless of program complexity.
The results suggest that verified software engineering is advancing faster than anticipated, potentially becoming more scalable than unverified methods in AI-assisted programming.
The project involved translating approximately 6k lines of Rocq into 25k lines of Lean, with AI autonomously handling 97% of the statements.
Human intervention was required for only 6 extreme-difficulty statements, totaling ~15 hours of effort.
The work highlights the potential for automated verification to address the bottleneck of human code review in AI-generated software.

Hasty Briefsbeta