VLA — arithmetic that carries its own proof.
Floating-point silently accumulates error; the same code gives different answers on different hardware. VLA replaces that with exact, error-free arithmetic — bit-identical on any GPU, at production speed, and every result ships with a reproducible SHA-256 receipt.
Personal tooling I built & run — proprietary, used across my own projects. Shown here as proof.faster than arbitrary-precision mpmath at the largest tested size (512×512), scaling with matrix size — and exact, not approximate
NumPy returns the wrong sign on Hilbert-matrix determinants; VLA is exact
same checksum on an RTX 4070 and a Tesla T4 — verified on Kaggle
a third party can reproduce any result byte-for-byte
How it works
Error-free transformations
Built on the classical results of Dekker and Knuth (two-sum, two-product) — every operation captures the rounding error exactly, so nothing is ever lost.
At GPU speed
110+ exact operators across linear algebra, transcendentals, and neural-net primitives — designed to run on the GPU, not a slow arbitrary-precision library.
A reproducible receipt
A prover/verifier protocol emits a SHA-256 receipt: anyone, on any hardware, can re-run the computation and confirm the result is identical to the bit.
Where it's been proven
The Wilkinson polynomial
A classic textbook example (Wilkinson, 1963) — double precision gets 4+ roots wrong; VLA gets all 20 exactly.
10,000 quantum gates
Unitarity (U†U = I) preserved exactly across 10K operations — a verified classical baseline for quantum work.
Interactive benchmarks
VLA vs mpmath across matrix sizes, the NumPy Hilbert failure table, and cross-GPU checksums.
Internal tooling I built and run
VLA is proprietary infrastructure I use across my own projects — it powers exact computation in Neruva and more. Shown here as proof of the work; happy to talk roles or research.