Run it twice. Get the same answer.
Determinism is the first pillar: the verify in "verify the law, read the reason." A model whose computation is governed by laws you can check, not a result you have to take on trust.
Most inference is quietly non-deterministic.
GPUs reorder floating-point work for speed, so the same prompt on the same model can return subtly different numbers, and sometimes different words, run to run. Most stacks accept that. We didn't.
On a fixed prompt, our kernel returns byte-for-byte identical token IDs, every single run. A fixed compute order, deterministic math, and a parity test on every change. Reproducibility isn't a nice-to-have here. It's the whole point.
From scratch, in pure Mojo.
The kernel is written from the ground up in pure Mojo, and the hot paths stay pure Mojo by policy, so there's no black box in the loop we can't account for. It runs at NVFP4 precision on Blackwell.
And the stack isn't locked to one vendor: it runs on NVIDIA Blackwell today, with support for AMD coming soon on consumer hardware. Your model, your silicon.
Here's the part we're proud of: on the newest hardware, our from-scratch kernel runs clean through a class of Blackwell bugs we found and fixed ourselves, and we're contributing those fixes back upstream. A small team, our own kernel, doing real systems work on some of the newest silicon in the world.
A quarter of the bits. All of the signal.
Lawful by construction.
Deterministic math, a fixed compute order, and byte-parity tests on every change. Same input, same output, provably. That reproducibility is what makes the next thing possible: a pipeline that can check its own answers instead of hoping.
You can't verify what changes every time you look.
A model shouldn't be a law unto itself. It should be governed by laws, and the first is that it holds together: a computation that contradicts itself, run to run, can't be checked against anything, inside or out.
Determinism is that first law made real: pin it down, reproduce it, audit it. And it's the ground the second condition stands on: that the reasoning also holds up against the world. Self-consistent within (Nomos), true to the world without (Logos).
A model shouldn't be a law unto itself — it should be governed by laws. The first: that it never contradicts itself.— The first pillar