Engineering deep-dives, research, and benchmarks you can re-run. Written by the team building Derivon.
Automatic differentiation is the right tool for first-order gradients and the wrong tool for iterated high-order operators. Here is the regime where a closed-form derivative tower wins decisively - and why.
A drop-in, bit-identical replacement for the default autograd Laplacian in neural-network variational Monte Carlo - with the memory curve flattened.
Speed is necessary but not sufficient for scientific trust. Here is how Derivon attaches falsifiable, machine-checkable certificates to a computation.
Soft penalty terms make PINN training ill-conditioned. Structural cages enforce incompressibility and conservation exactly - and keep the closed-form fast path end to end.
We publish a CPU smoke tier that runs in seconds and a GPU headline tier transcribed verbatim. Here is how we keep performance claims honest and reproducible.