Setun Labs goes online

Setun Labs opens its doors, and this site is its front door. We are a small independent lab building AI that cannot lie about the domain it serves, not because it is larger than the competition, but because its architecture makes correctness unavoidable.

What we stand for. Three legs.

Mechanism before scale. We do not train larger models and hope they stop hallucinating. We design deterministic structures, verification loops, constrained proof grammars, ternary state machines, that bound what the model can say and check what it does. If a 1-billion-parameter specialist with the right structure outreasons a 100-billion-parameter generalist without it, that is not surprising. That is the point.

Refutation over abandonment. A negative result is a coordinate, not an exit sign. When a test returns [KILL], we do not pivot, rebrand, or declare the field saturated. We map the wall, redesign the mechanism, and rerun the protocol on the same frozen sample with the same pre-registered thresholds. The only terminal state is understanding.

One domain, zero doubt. We build specialists that do one thing perfectly. Every output is either verified correct by the symbolic engine, or explicitly flagged as uncertain. General capability is the enemy of guaranteed accuracy. We would rather build a math model that cannot lie about proofs than a generalist that is occasionally right about everything.

What we're working toward. Three arms, all in development.

Models: the Grimm Series, open, ternary-weight specialist models built on the Static S-Cube architecture (three interacting channels per weight instead of sixteen bits of noise). The first is a math specialist, currently in pre-training proof. It is not the product. It is the evidence that our method produces verified accuracy at any scale. When it proves itself, it unlocks the series: physics, code, chemistry, formal logic, each running offline on hardware you already own.

Apps: Pocket AI Hub (working title), one app to run those models on your phone and recruit your computer for the heavy jobs. No accounts. No data centers. No one watching over your shoulder.

Tools: the open-source plumbing underneath both, the Battery evaluation framework, the grammar engine, the verification stack, free for anyone to inspect, fork, and refute.

The honest status. Nothing on this site is released, for sale, or downloadable yet. Every project here is being worked on right now, and each one publishes only when it is real and measured. Our pre-training uncertainty audit (Doc 23) returned [KILL] on its first configuration, grammar too narrow, oracle too low, so we are expanding the typed proposal space and rerunning the same frozen protocol before any GPU training begins. We publish that too. You should know exactly where things stand before you give us your attention, and especially before you give us your support.

How to stand with us. If this is the AI future you want, verified, owned, unstealable, here is how you can help, in order of what we need most.

Investors. We are not raising a round yet. We are building the artifact that makes a round possible: a working 1B math specialist with verified accuracy on public benchmarks. If you invest in deep technical teams before the pitch deck is polished, and you have the patience for six months of engineering before a term sheet, we want to talk. Send us what you have backed before and what you look for in a first check.

Grants. We are a better fit for program officers who fund mechanism research than product demos. If your program supports neuro-symbolic AI, verified computation, or hardware-efficient architectures, we have a pre-registered evaluation framework and a reproducible audit trail ready for your reviewers. We can submit a full technical brief on 48 hours notice.

Collaborators. We need three things from the outside: formal-methods reviewers who can stress-test our proof engine, hardware engineers who have shipped quantized kernels on ARM and CUDA, and domain specialists in math education who can tell us whether a verified proof matters to a student. If that is you, we will make time.

Sponsors and labs. We are not a closed shop. If you run a lab, academic, corporate, or independent, and your work touches verified AI, neuro-symbolic architectures, efficient inference, or specialist reasoning, we will happily collaborate. Joint papers, shared benchmarks, co-designed protocols, hardware sponsorships, compute grants, we are open to any arrangement that preserves our independence and our pre-registration standards. We do not take exclusive deals, and we do not embargo negative results. Everything else is on the table.

Users. Sign up for the list. When the first Grimm model is real and measured, you will know before the press release. If you are an educator, a researcher, or an engineer who needs math or code reasoning that cannot fail silently, tell us your use case. We build for specific people, not abstract markets.

Everyone else. Share this page with the one person you know who cares more about correctness than hype. That referral is worth more than a tweet.

Write to us at setunlabs@gmail.com. We read every message, and we reply to the ones that show us something we did not expect.