About

What is The Numerical Proving Ground?

The Numerical Proving Ground (TNPG) is our research program. It aims to assess the reliability of numerical software against its specifications, no matter who or what wrote it, beginning with numerical linear algebra kernels.

We pursue this aim through two projects. VSCL is a Lean library of numerical analysis whose proofs Lean checks. VERNA is designed to assess kernels against explicit specifications, using formalized results from VSCL to guide the choice of checks and inputs.

This site gathers our paper, our records, and our projects.

What would Turing do?

In our paper, we draw on Turing's work, including Rounding-off errors in matrix processes (1948) and Checking a large routine (1949), to set six design goals. They guide which results we formalize, how we review them, and how we use them to assess numerical software.

Together, the goals describe a program for developing reusable, checkable numerical guarantees as algorithms, representations, and precision choices change. Our ambition is to combine numerical expertise with AI-assisted exploration and formal proof checking, so that software generated faster comes with better evidence of its reliability.

Turing analyzed rounding errors and proposed a way to check programs when digital computers were new. With today's proof checkers and AI systems, he might well have started such a program himself.