sturnia

A manifesto for shared intelligence

The Sturnia
Manifesto.

Intelligence we all own.

It starts with your agent. It grows with your community. Together, something greater.

Read the story
Your agent. Your co-op. A federation of communities.

The story

The wire

Eighty years ago, the power companies looked at rural America and decided the wires weren't worth stringing. Not enough profit in lighting a farmhouse.

So the farmers strung the wires themselves. They formed cooperatives, neighbor-owned, neighbor-governed, bought the poles and copper with their own money, and lit their own towns. Within a generation those co-ops had electrified more of the countryside than the companies had ever planned to. Hundreds of them are still running today.

Eighty years later, the same math is being run on intelligence. And once again, there are people the wires don't reach.

Today's farmhouse is a therapist's office. It's a clinic, a law practice, a union hall, a contractor with twenty years of fixed boilers in his filing cabinet. The people whose work would benefit most from AI are exactly the people who can't paste that work into a chatbot, patients, clients, cases, contracts. For everyone else, AI is cheap and everywhere. For them, it never arrived.

The farmers didn't wait, and they didn't smash the generators. They built their own.

So will we.

Individual homes become a shared network. The people who use it own it.

What Sturnia is

Three circles

Your agent

The AI that's allowed to see everything.

It runs on your computer, sized to your hardware, and works with no network at all. Your files, your notes, your mail, it reads what you trust it with, and by default nothing leaves the machine. Local-only mode makes that checkable by anyone: watch the network connection and it stays silent, and we publish the test so you don't have to take our word for it. When something would leave, a question routed to a specialist, you see exactly what's in the envelope before it goes. No silent pipes.

Your files, your agent, your answer. Private work stays on your machine by default.

Your co-op

A community that pools its machines and its knowledge.

A clinic network, a trade association, a campus lab, a homelab club, or a street of neighbors pooling machines behind a specialist community they care about. One or two serious, always-on machines anchor it: the generator shed. Everyone else's devices pitch in on work that survives a lid closing, searching, drafting, checking, critiquing. We never split one model's brain across the neighborhood; that graveyard is full. But the co-op's real treasure isn't compute. It's the corpus: the manuals, the resolved cases, the donated examples that only this community has. That's what turns a model into a specialist.

Members contribute idle compute and curated knowledge. Always-on hosts keep the co-op available.

The federation

Specialists that answer each other.

Here's how a federated answer is designed to work: co-ops publish what they're good at, and a hard question travels to the two or three that actually know the field. Their answers come back checked against sources, cross-examined, signed, with disagreements shown rather than smoothed over. No company in the middle.

One plain rule, stated up front: anything you approve for sending is seen by the co-ops that answer it. The preview exists so you decide, strip the names, generalize the details, or don't send at all. Truly private work never needs to leave local-only mode.

And the honest hard part, named: trust between strangers' co-ops, fake identities, poisoned specialists, gamed reputations. Our published design review covers all three, and the Phase 4 gate is that our scheme survives public, bountied attack before any reward attaches to anything.

An approved question reaches specialist co-ops. Their sourced answers come back to you.

A murmuration: many small things, no leader, moving as one mind.

Slow is smart

We don't race datacenters. Racing is for chatbots.

Ask Sturnia a hard question and, by design, it does what no chatbot does: sends it to the specialists who know the field, lets them answer independently, makes them argue, checks the claims against sources, and returns one answer with the reasoning, and the dissent, attached. It may take ten minutes. It will be worth ten minutes.

The best answer your network can produce, by lunch. That's the product. Depth costs more than speed, and it should: quick answers are cheap; answers that survive review cost what rigor costs.

Independent answers are compared, then checked. The result includes sources and disagreement.

The deal

What you give and what you get, and it holds even if you don't care about a single ideal on this page.

You give
  • idle cycles on machines you already own
  • examples you explicitly choose to donate (revocable, deleted from the corpus and excluded from all future training; anything already trained on them can only be unlearned by retraining or rollback, which is exactly why publishing is gated behind privacy budgets)
  • your judgment, ratings, corrections, hard questions.
You get
  • an agent that actually knows your work, because it's allowed to see it
  • once co-ops federate, access to every specialist your federation has, from HVAC troubleshooting to case law
  • and, as the network matures through the gates below, a local AI that inherits what it learns: the specialists it can call, the answers it can trust, the knowledge it plugs in.

Your AI knows what the network knows. Intelligence that compounds for everyone who joins.

Contribute compute, knowledge, or judgment. Get access to the intelligence your community builds.

Principles

  1. Useful alone, better together.

    Everything works on day one with zero network. The network is a multiplier, never a requirement.

  2. You see everything that leaves.

    Raw data stays on your machine. Anything outbound is shown to you first. Anything the network publishes that learned from members is trained under a published, mathematically enforced privacy budget (the technique is called differential privacy), strong, measured protection, honestly bounded.

  3. Smart over fast.

    Depth, verification, and debate over milliseconds.

  4. Nothing is earned but verified work.

    What cannot be verified cannot be rewarded.

  5. No single kill switch, including ours.

    If the founders vanish, Sturnia keeps running. One deliberate exception: new model versions go to a small volunteer group first, and many independent hands hold the rollback, because unreviewable weights shipped to everyone would be a kill switch too.

Private files stay local by default. Review the exact outgoing question before approving it.

What Sturnia is not

  • Not a coin.

    There is no token, and none is planned. Inside the federation, co-ops keep score with plain non-transferable credits, bookkeeping, not currency. They cannot be sold, transferred, or converted into anything tradable, not now, not later. If that ever changed, it would be a different project. Anyone selling you a "Sturnia token" is running a scam.

  • Not a ChatGPT competitor.

    If the cloud already serves you, use it, we do. Sturnia exists for the work the cloud can't be allowed to see, and for the question a lone model can't answer as well as a network of specialists.

  • Not anti-AI.

    We're pro-ownership. Nobody strings their own wire because they hate the light.

  • Not a promise dressed as a product.

    See "Where this stands," below. We publish our own failure conditions.

Where this stands

Honesty, since we're asking for your trust: today, Sturnia is a design and a brutal published review of that design. The software comes next, in the open, gate by gate.

We recruited our harshest critics first, a 36-finding design review plus an independent external model review, and published everything, including what broke: our first architecture had real flaws, and the review rewrote half of it. You can read the before and after in the open repo.

Every phase below ends in a gate: a test that must pass, measured and published, or we stop and say so. No gate, no next phase. That's the deal we're making with you before we ask for anything.

Roadmap

  1. Phase 0

    The blueprint

    ✓ (done, save the trademark check)

    Mission, manifesto, architecture, and a public design review that rewrote half of it. Shipped: the documents you're reading.

  2. Phase 1

    The agent

    (next)

    A local evidence assistant for people whose work can't touch the cloud: reads your documents, answers with citations it has verified against your sources, remembers what you teach it, and proves that local-only mode sends nothing anywhere.

    Gate real users solving real work measurably faster, and beating the same model without Sturnia's workflow on a published test of ~200 real tasks. If it can't, we don't ship it.

  3. Phase 2

    The first co-op

    One real community, a clinic network, a contractors' association, a campus lab, pools its machines and its corpus into the first specialist. Anchor hardware hosts the model; everyone's devices share the work that tolerates interruption.

    Gate the co-op must beat the best single workstation in the group on verified answer quality, throughput, and cost, measured across four configurations, published. If pooling loses to one good machine, we say so and redesign.

  4. Phase 3

    The first ten

    Hand-picked co-ops with complementary specialties, connected: capability profiles, routing, ensemble answers with recorded disagreement, and plain credit bookkeeping between co-ops. Early co-ops shape the standards everyone else inherits, first pick of specialties, a permanent seat in eval custody, priority in routing while they hold the quality bar. Before Phase 3 ships, we'll publish exactly how early standing works: in governance, not in anything that trades.

  5. Phase 4

    The learning loop

    The network starts teaching itself: specialist adapters shared across the federation, a library your local agent inherits, and the rules that decide which contributions count, built in the open, attacked in the open, with bounties for breaking them. Rewards attach only after the attacks fail.

  6. Phase 5

    There is no Phase 5 until Phase 4 survives its adversaries.

FAQ

What can I actually use today?

Nothing yet. The blueprint is done and public (Phase 0); Phase 1, the local agent, is next. What exists today is the design, its published review, and the repo. The first thing we ship will be useful with zero network on day one.

Why not just use ChatGPT?

If you can, do. Sturnia is for the work you can't send to the cloud, and later, for questions that deserve a network of specialists instead of one generalist. If nothing in your life is too private for a datacenter, we're not your tool, yet.

Is my data safe?

Safer than a pinky-swear: it's designed into the architecture and testable once shipped. Your data stays on your machine; local-only mode will be verifiable by anyone, watch the network connection, it stays silent, and we publish the test. Anything outbound is previewed by you first. Donated examples are explicit choices, revocable as defined above. And anything the network publishes that learned from members is trained under formal privacy budgets with the numbers published. The honest limit: a specialist trained on a community's donated examples encodes what it was taught, which is why donations are previewed, why publishing is gated behind privacy budgets, and why we treat this as engineering, not fine print.

Can I turn privacy on and off?

Yes, it's a switch, and it's yours. Privacy on is local-only mode: your agent uses only the model on your machine, nothing routes to the network, and you can check the silence yourself, watch the connection, we publish the test. Privacy off allows routing: a hard question may travel to the specialists who know the field, and you see exactly what's in the envelope before it goes. Flip it per question or leave it on for good. And the honest note: until co-ops federate, there is nowhere to route to, today every answer is local, whichever way the switch points.

Does Sturnia make my practice HIPAA- or GDPR-compliant?

No tool does that on its own, compliance lives in your practices, not in software. What the architecture removes is the biggest structural risk: in local-only mode no third party ever holds your clients' data, because nothing leaves your machine. Your professional obligations stay yours, and Sturnia's answers are a starting point for your judgment, never a substitute for it. Compliance documentation ships with Phase 1; until then, assume nothing and ask your own counsel.

Is there a token? Will there be?

No, and none is planned. Non-transferable credits keep score between co-ops; they cannot be sold, transferred, or converted into anything tradable, not now, not later. We wrote this into our founding documents on purpose.

Who pays for this?

Today, nobody can give us money and we're not raising. Inside a federation, co-ops balance work with non-transferable credits. How the steward foundation gets funded is a Phase 3 decision, and we'll publish it before it matters.

What hardware do I need?

To run your agent: the computer you already have, the app sizes its model to your machine (current target: usable from ~8 GB of RAM). To anchor a co-op: one serious box that stays on, think 64 GB+ unified memory or a 24 GB GPU. To be a member of a co-op: any device; the work assigned to it tolerates sleep.

How do I start a co-op?

The first ten are recruited by hand, a real community, a real corpus, one anchor machine, and one person who cares. If that's you, write to us. After the first ten prove the model, the playbook becomes public: technical and legal, entity templates included, because an unincorporated co-op shouldn't carry personal liability. (Templates are a head start, not legal advice, budget an hour with a local lawyer.)

Who owns Sturnia?

The code is open source. The shared assets, base model builds, marks, eval custody, will be held by a steward foundation, not a company. Co-ops own themselves. That's the credit-union model, and it's a hundred years old for good reasons.

Why "Sturnia"?

From Sturnus, the starling genus, thousands of starlings turning as one murmuration, no bird in charge. It's the architecture, as a word.

The invitation

If you have a computer that sleeps at night, you have something to contribute.

If you know things, a trade, a field, a place, your knowledge could seed a specialist.

If you have a community, you have the makings of a co-op.

And if your work is too private for the cloud, patients, clients, cases, contracts, you are not an edge case. You are who this is for.

The starlings don't ask permission.

Eighty years ago, the answer was to string the wire yourself.

It still is.