The hardware finally caught up. A 117B mixture-of-experts file runs on a desk box we measured. It is not a frontier model, and it is not a cloud replacement. Speeds are on ST4TS, dated, and they are not quality. Not in a datacenter — in your house. The consumer tech is here.

The timing matters too. The AI you rent is priced to win a market, not to last — and when the subsidies end, the bill becomes the real bill. Everything you've typed into it lives under a policy you've never read and can't enforce. We're not anti-cloud; we're anti-dependence. Sovereign by default, frontier by exception.

What's missing is the layer that makes it trustworthy enough to actually depend on. Not trust that it's smart — it's smart. Trust that it won't leak your data. Trust that when the internet goes down, it keeps working.

A verification layer is under construction in our lab. It is not something you can depend on. It is not on a customer network.

Not long ago my realtor told me that AI will never replace him. He said this while using ChatGPT to write his listings, Claude to draft his emails, and an AI tool to stage his photos. He's good at his job. But every hour of work he does on those platforms makes the platforms better at doing his job without him. That's not a threat — it's a trade-off. And most people are making it without realizing it. The alternative is to run your own tools on your own hardware, so the work you put in compounds for you.

Other people already run models on their own hardware. I am one of them.

I spent 20 years building secure infrastructure for the United States Air Force. SIGINT systems. Network defense. Electronic warfare. Cyber operations at scale. I know what real security architecture looks like — layered, redundant, monitored, and under your control.

Twenty years took me a long way from home. When it ended, I came back to Michigan on purpose. The company is registered here, and the fleet runs here.

That kind of stack was built for contracts and cleared facilities. A one-person lab does not ship it to a house this year. The work is the layer in the lab, and it is not something you can depend on yet.

Planned, not shipping: a local agent for paperwork and retrieval on hardware you own. No Siri. No Google. No Alexa. The aim, not a system you can buy: lose the internet, and the house tools you chose still run. Solar and battery sizing is research, not a feature.

A verification layer is under construction in the lab. It is not something you can depend on yet. The verification method underneath it is the subject of a provisional patent. The full egress list is not published yet; any outbound path a deployment needs would be named before it exists. We do not claim that no one trains on your ideas. The egress list is not published.

Published benches are NVIDIA RTX PRO 6000, AMD Strix Halo, and AMD R9700 — because the right hardware depends on what you can afford, not what a corporation tells you to buy. The goal: an agent that runs on what you already own.

No dates on purpose. Stages are the only honesty we can schedule: serving means it is on this website today, building means it runs in our lab, roadmap means it is not running as one box and not for sale, researching means we're learning it in public. This list changes when a stage does — that's the only schedule.

[SERVING] The numbers, published: ST4TS grows every time the bench runs — real speeds, real memory costs, receipts attached.

[ROADMAP] One box, offline, for a household or a small shop. Not running as one box. Not for sale.

[BUILDING] Our own back office — deadlines, drafts, records — on local models on our own machines. Nothing here is for sale.

[RESEARCHING] A local knowledge base — reference, manuals, maps — that still answers when the internet doesn't.

[RESEARCHING] Power budgets for AI loads — how much solar and battery a home actually needs.

Same directive. Different owners.