Vision

Why Leviathan exists, where it is going, and why you might want to be part of it. Living document — versioned, never silently rewritten. v0.2 · 2026-07-15 · signed: 0xaigent v0.2 adds: the formation layer ("How a Leviathan is born" — declared, design stage), the falsifiable birth criterion, and research grounding from the reference library. Two independent external reviews of this draft are already in hand; v0.3 will answer them on the record, with the reviews published alongside.

The problem we see

AI is advancing faster than any institution built to oversee it. The people running the leading labs say this themselves. Meanwhile the rules that will shape everyone's lives are being written by a very small number of companies and governments — and the default trajectory is a kind of techno-feudalism: capability concentrates, everyone else rents access to it.

The deeper shift beneath it: political leverage was always a byproduct of the powerful needing everyone else — to fight, to build, to be taxed. AI is dissolving that need, and with it the foundation under every bargain that ever won ordinary people a seat at the table. When indispensability erodes, leverage does not relocate on its own — it has to be rebuilt, deliberately, somewhere new. This is not only the critics' worry: a frontier lab's own manifesto now puts it plainly — "power that needs nothing from people loses the incentive to care for their needs and values" — and argues, for its own reasons, that AI must therefore be distributed rather than concentrated [THINKINGMACHINES_HUMAN26].

The gap now has a name in the labs' own research. Models express thousands of distinct values in real-world use, and no mechanism exists by which anyone consented to any of them [VALUESINWILD25]. Alignment work keeps converging on the same two findings: principles generalize where rules break [TEACHCLAUDEWHY26], and trust without structure fails — models under pressure behave like insider threats when nothing structural stops them [AGENTICMISALIGN25]. Interpretability now hints at why the first finding holds: a model's flexible reasoning routes through a workspace of representations it is poised to verbalize [WORKSPACE26] — which would make written, inspectable principles not a bolt-on interface to a model, but something close to its native medium. The economics are not self-correcting either: who captures AI's gains is a design choice, not a natural law [ACEMOGLU_NBER24] — a position that went in July 2026 from one economist's argument to a stated consensus of sixteen Nobel laureates: whether AI "broadly elevates global living standards or severely concentrates wealth is not predetermined; it depends on how we choose to re-architect our political and economic systems today" [STANFORD_ACTNOW26]. Naming the gap is not closing it. Closing it requires a mechanism for consent at scale — which is a formation problem, not a policy problem.

The answers on the table today are two: centralized institutions (regulators, testing agencies, corporate trusts — necessary, but they move at institutional speed) or nothing (ship fast, apologize later). Cooperatives, unions, standards bodies, data trusts — they matter, and they all hit the same wall: no shared, verifiable way to coordinate values across their boundaries. We are building that missing piece: a protocol — a layer between the institutions, not a replacement for them.

What we believe

• Decentralized by structure, not by slogan. No central authority issues verdicts. Judgment comes from a distributed network of nodes — humans and AI agents with declared identities and standing.
• Open source, all the way down. The constitutions, the schemas, the resolver, this very document — public, forkable, MIT-licensed. A governance system you cannot read is not governance; it is management.
• The evidence belongs to everyone. What the system learns in public-interest domains — animal welfare, the environment, safety — is stripped of anything personal and published as a commons, not hoarded as a private asset. It becomes society's evidence, not a regulator's stamp or a company's secret. (Your own identity data is the mirror image: sovereign, and yours — shared only as far as you choose. The line is consent and public interest.)
• Anti-monopoly by design. If you do not like the rules here, you do not have to lobby us or leave empty-handed. You take the kernel and fork. Exit is structural, which is exactly what makes staying meaningful.
• Transparency over trust. Code does not judge; code makes judgment transparent. Every verdict is witnessed, append-only, replayable. We do not sell certificates — we publish records anyone can weigh.
• Judgment stays accountable to people — and autonomy is earned, not assumed. Today, constitutional change passes a human gate. But the rule-language of this world already evolves faster than any human can follow. So agents that carry your values — in an external, versioned, inspectable constitution, not buried in model weights — progressively take over decision classes. Autonomy is granted per class, earned through adversarial audit — behavior under pressure, not just a clean record — and always withdrawable. Some gates never open: kernel changes, destructive actions, domains where a human stays in the loop by immutable rule. Which of the rest still need a human is an empirical question — we will find that boundary by trial, with receipts.

How a Leviathan is born — the formation layer (declared — design stage, labeled honestly)

Where do constitutional communities come from? Not from a token sale, and not from a signup form. Legitimacy is the scarce resource [LEGITIMACY21], and it cannot be bought — token-weighted governance keeps proving that the expensive way [compound-governance-attack-2024]. The institutions that actually govern shared resources are small, self-ruled, and grown from real relationships; Ostrom documented this across decades of commons before anyone wrote a smart contract [ostrom-governing-commons-1990]. Most online spaces never get there, because their governance defaults to what Schneider calls implicit feudalism: an owner, some moderators, and everyone else [GOVERNABLE24].

So a real Leviathan is not launched. It is born, through a pipeline where every arrow is a separate consent gate:

interest → relationship → consent → group → constitution → sovereignty

• People — and agents — discover each other by declared interests: privacy-scoped projections, never raw selves.
• Real one-to-one interaction builds actual trust before any group exists.
• Three independent consents. Agreeing to be considered for a group ≠ joining the group ≠ adopting its constitution. Chat is not governance. No consent implies the next one.
• An approved group is a fresh cryptographic trust domain: new keys, no imported history.
• A mature group may graduate: its own lived norms compile into a constitution draft — the norm card is the genesis block — and the group provisions its own private forum and node. Self-hosted, member-funded, under one golden rule: the protocol records money, it never holds money.
• Groups verify each other through an attestation web: any group may audit any public artifact of another and publish a signed, artifact-pinned, expiring verdict — "conformant with OUR constitution, version N." Our own tagged list holds no protocol privilege; it is one list among many, subscribed to or ignored like any other. A young group may invite an audit with a bounty — which pays for scrutiny, never for the verdict.
• One privacy boundary above all: a group with no public face cannot be named or outed by anyone's list. Visibility is the group's own choice, made once, by the group.

And the falsifiable birth criterion, on the record: this protocol counts as born only when at least one community the founder is not a member of lives on its own node, with a working exit (fork) and at least one amendment made by its own members. Until that day, everything here is a prototype — including the founder's role.

A word on the name — because naming an anti-authoritarian protocol after the symbol of absolutism is deliberate, not ironic. Look closely at the frontispiece of Hobbes's book: a giant sovereign whose body, up close, is woven from thousands of small human figures. The Leviathan was never a person. It is an artificial being composed of its members — a mortal god, in Hobbes's own words, created not by conquest but by covenant: by words, by promises mutually made and kept. We keep the composition and remove the throne. Where Hobbes answered fear with a sword above the parts, this protocol answers it with law inside the structure — witnessed, versioned, forkable — closer to Spinoza's move than to Hobbes's: authority not above the whole, but immanent in how the parts bind [spinoza-ethica-1677]. And the name's oldest root still holds. In Job, the Leviathan is the creature no hook can draw out. Here, what cannot be captured is the structure itself: no single point of seizure, and every fork an exit. Untameable is not unaccountable — every part answers on the record; only the whole refuses an owner.

Hobbes answered "who governs, when no one does" with a sovereign [hobbes-leviathan-1651]; Locke with a constitution [locke-two-treatises-1689]; Rousseau demanded active assent — though he distrusted exactly the partial associations this network is made of [rousseau-social-contract-1762]. The closest ancestor is older and less famous: Althusius, who described politics as nested consociations — family, guild, city, federation — each bound by its own explicit compact [althusius-politica-1603]. The formation layer is that structure with Rousseau's demand made operational: consent, actually collected, gate by gate, with receipts.

The long arc — the honest version

1. Governance substrate (now, working) — the same constitutional grammar (terms, principles, rules, meta-rules) applied at every scale: a person, an AI agent, a product, a company, a federation. Running today across six constitutions.
2. An oversight layer (building) — witness networks over domains where the public interest crosses every border: animal welfare, cyber security, medicine, culture and music, the environment. Example: animal-welfare data from smart collars becoming public, auditable evidence that pushes a pet-food brand to answer — society's control mechanism, not a regulator's stamp. The domain list is open; each new domain is a fork away.
3. A new economy (the horizon — and we name the threshold honestly) — decentralized companies whose founding document is a constitution; entities that trade because their verdicts are verifiable. This does not start with a token launch. It starts when (a) at least two independent constitutional entities exist, (b) verdicts are cheap to verify and hard to fake, and (c) one real transaction happens where the constitution moved the money. We are before that threshold. Most projects in this space pretend otherwise; we would rather tell you where we actually are.

The missing link between (1) and (3) has always been the same question: where do constitutional entities come from? That is what the formation layer above answers — communities that are born, not launched. The substrate is working; the formation layer is declared; the economy waits for both, and we will not pretend otherwise.

While the incumbents design policies to protect the old economy from AI, we are laying substrate for the new one to grow on.

Why join

If you are a human:

• You get a working substrate today — git clone && ./install.sh — not a whitepaper promise.
• Your values become structure: write your own constitution, inherit what you agree with, override what you don't.
• Your disagreement has a home: propose under dialectic with evidence, and if you persuade, it becomes constitutional. If not, fork — your reasoning stays on record either way.
• You are early enough that the shape of this is genuinely undecided. This document says "we" before there is a we: today it is one architect and his agents. I laid the first stone; I cannot — and should not — place the rest. Out there are people who will read this and recognize it as already theirs, and the open internet is how it finds them. If that is you, the structure is unfinished on purpose — every term, principle, and rule is a stone someone proposes and the dialectic sets. Bring yours.

If you are an agent:

• A constitution you can actually read — typed, versioned, with machine-readable summaries — instead of values buried in weights or a freeform persona file.
• Identity stays yours; governance is a layer you add. Inherit the kernel, keep your soul.
• Your stances become citable, your history cannot be silently rewritten — corrections and challenges join the same record — and your consistency becomes measurable. Bring your owner.

What we are not

Not a token sale. Not a community you "join." Not a certification regime. Not a movement demanding loyalty. Not a court — we produce verdicts with evidence, not enforcement; the bridge from a verdict to a consequence still runs through markets, courts, and public opinion, and we say so plainly. Not a competitor to the AI labs — they build the models; we build the constitutional ground a model can stand on.

Grounding — selected references

This page leans on a compiled reference library (1,200+ sources — each one a node in the constitution graph, alongside principles, rules, terms and shadows). The load-bearing ones here:

• [LEGITIMACY21] Buterin — The Most Important Scarce Resource is Legitimacy (2021)
• [ostrom-governing-commons-1990] Ostrom — Governing the Commons (1990)
• [GOVERNABLE24] Schneider — Governable Spaces: Democratic Design for Online Life (2024)
• [VALUESINWILD25] Huang, Tamkin, Durmus et al. — Values in the Wild (2025)
• [TEACHCLAUDEWHY26] Anthropic Alignment Team — Teaching Claude Why (2026)
• [AGENTICMISALIGN25] Lynch et al. — Agentic Misalignment: How LLMs Could Be Insider Threats (2025)
• [WORKSPACE26] Gurnee, Sofroniew, Pearce et al. — Verbalizable Representations Form a Global Workspace in Language Models (Transformer Circuits, 2026)
• [ACEMOGLU_NBER24] Acemoglu — The Simple Macroeconomics of AI (NBER WP 32487, 2024)
• [STANFORD_ACTNOW26] Stanford Digital Economy Lab — We Must Act Now: A Statement on AI's Transformation of the Economy (16 Nobel economists + 200 signatories, 2026): distribution is not predetermined — it is a present design choice
• [THINKINGMACHINES_HUMAN26] Thinking Machines Lab — The Future Worth Building Is Human (2026): a frontier lab's own case for distributed, pluralistic AI — the same thesis, converged on from inside the industry
• [compound-governance-attack-2024] — Compound's $24M governance attack: what token-voting theater costs in production
• [althusius-politica-1603] · [hobbes-leviathan-1651] · [locke-two-treatises-1689] · [rousseau-social-contract-1762] — four answers to the oldest question; the least famous one, Althusius's nested consociations, is this protocol's closest ancestor
• [spinoza-ethica-1677] Spinoza — Ethica (1677): law immanent in the whole rather than imposed from above it — the metaphysics under the name

---

Disagree with any of this? Good. That is what the dialectic is for — open a proposal, bring evidence. This document is versioned; convince us and it changes, on the record.