The Numerical Proving Ground

The Numerical Proving Ground is a research program that aims to assess the reliability of numerical software. This site gathers our machine-checked proofs, records, and notes.

AI systems now write numerical software far faster than experts can review it, and for more than five decades expert authorship and review have been the main assurance that such software can be trusted. When review cannot keep up, further evidence is needed that the software meets its stated requirements, such as accuracy within a specified tolerance.

Tests can expose failures, but they cannot show that none remain. A proved error bound covers every input that satisfies its assumptions. Machine-checking such bounds for more algorithms needs larger libraries of formalized numerical analysis, and the libraries surveyed in our paper hold very little of it.

A case in point

In 2026 the GPU MODE community held public competitions for fast linear algebra kernels (a kernel is a routine for one numerical task, such as factoring a matrix), open to participants and their AI agents. The winner of the second competition on QR factorization, a standard matrix factorization, ran 107 times faster than the baseline, which calls NVIDIA's QR routine through PyTorch, and passed every accuracy test that the competition ran.

We then ran each of that competition's five highest-ranked kernels 1,050 more times, on new inputs from the competition's own generator, judged by the competition's own checker. The winner failed 155 of these runs, the other four failed 46, 327, 3 and 357, and the baseline passed all 1,050. Passing a fixed set of checks does not establish a kernel's accuracy.

The record of our audit holds its data for twenty kernels from four competitions. Eighteen failed on further inputs: all ten QR kernels, all five Cholesky kernels, and three of the five eigendecomposition kernels.

Our two projects are meant to supply evidence of reliability. VSCL exists today. Our paper states the goal of VERNA and leaves its design to a separate publication.

VSCL

A Lean library of numerical analysis, written and reviewed by AI agents and checked by Lean.

2026-11-16
VERNA

VERNA is the Verification Engine for Reliable Numerical Algorithms. It is designed to assess numerical linear algebra kernels against their specifications.

under construction

VSCL, the Verified Scientific Computing Library, is our public library of numerical analysis in which Lean, a program that checks proofs, has checked every proof. Its catalogue lists the items of its sources that it has formalized and that its own records mark as reviewed.

Trustworthy Numerical Software in the Age of AI: What Would Turing Do?

Our paper, on arXiv. It presents our audit of a GPU QR competition, VSCL, the goal of VERNA, and six design goals that draw on Turing's work.

2026-11-16

A record holds the data behind a result, so that each stated result can be checked, and replayed where its inputs and code are public.

GPU MODE linear-algebra kernel audit, 2026: results record

The data of our audit of twenty kernels from four competitions: verdicts, descriptions of the inputs, hashes, and scripts.

2026-11-11
The RandNLA case study in VSCL: statements and proofs

The 47 statements that VSCL formalizes from the survey of randomized numerical linear algebra (RandNLA) by Drineas and Mahoney, with their Lean text, the 35 published proofs, and how many theorems of the library each proof depends on.

2026-11-16

Formalization work is growing quickly, and very little of it concerns numerical methods. A keyword search of the roughly 182,000 theorems and lemmas of Mathlib, the main library of Lean, finds 13 on numerical analysis. Lean libraries of numerical analysis have appeared only since the rise of AI agents: NLA-lib, a library of numerical linear algebra meant also for research done by agents, and our own VSCL. Related work shows these counts in two figures, and lists by topic the work that we build on and stand beside.

About