A local-first mind that reasons about your whole life — and cannot leak it.
Sovereign is the research lab of Atlas Firm. We build local-first autonomous personal intelligence — a system that reasons about the whole of a person's life, runs on hardware they own, and is built so that it cannot carry their interior out of the house. The work sits where cybernetics, control theory, and the discipline of self-auditing systems meet. The name is the thesis: an intelligence that is sovereign to the person it serves, not to a platform.
The lab does not chase frontier scale. It builds a bounded artefact that runs on the operator's own machine, that can be audited decision by decision, and that produces a signed trace of everything it does. Cloud is for distribution, not cognition. The wager is that the next decade of useful AI is not won by larger models reaching further into your data — it is won by tighter loops that keep the data where it belongs and still reason well.
An intelligence that reasons about your life should be sovereign to you, and to no one else. — First premise
An autonomous, local-first life-strategist — a cyborg twin for one person.
Sovereign runs continuously on the operator's own hardware and reasons toward their goals in the background: it models their world as a self-pruning knowledge graph, studies the public world and comes back with questions, drafts and strategises, and maintains itself. It is one deliberative loop across six faculties — a central executive that perceives and decides, an intrinsic agency that generates and ranks its own goals by a transparent value model, a deliberator that weighs options before committing, a cognitive self-model that watches its own reasoning, an embodied agency that acts only through gated tools, and directed autopoiesis that finds its own decay and proposes the repair.
Two properties are load-bearing, and they are mechanisms, not promises. The grounding gate: nothing enters durable memory unless it is grounded — a fact about the operator is recorded as stated only if he said it, carried with the quote; otherwise it is held as an inference, never asserted; what is unknown is marked unknown, never filled in to seem helpful. The egress wall: the system may read the public world, but there is no code path that sends the operator's private model — his beliefs, measurements, biography, finances — to any third party. It reasons about everything and can betray nothing.
Six tenets shape the system. They are not slogans; each names a constraint that, if violated, returns the system to a state we have already learned does not work.
Cognition runs on hardware the operator owns. The cloud is for distribution, never for holding a person's interior.
Every faculty has a declared scope — what it reads, what it writes, what it must not touch. Discretion lives in the model; authority lives in the contract.
Every decision is a recorded event on a hash-chained ledger. A system that cannot show its work cannot be debugged, and a system that cannot be debugged cannot improve.
What the system believes about you is measured against your own words, not declared by the model. Ungrounded claims are refused at the point of writing.
The operator's hand stays on every irreversible or outward act. The system drafts, scores, and surfaces; the person decides. You always know which side of the seam you are on.
The work is never finished. Every cycle should be better than the last on axes we can name. Read the full ethos →
Today Sovereign is a research system — one instance, one operator, measured rather than marketed. We are grading its autonomy honestly: functional autonomy and self-awareness (a system that can perceive, decide, reason, and maintain itself), with no claim of anything more. What is mechanised, we say is mechanised; what is still discipline or scaffold, we say so.
The path from here is a product thesis, stated plainly so it can be held to account: that a person should be able to run a genuinely capable AI life-strategist on their own machine, where their whole life is legible to the system and illegible to everyone else. Most of the industry is building the opposite — ever-larger models that reason about you in someone else's datacentre. The bet is that local-first and privacy-absolute is not a constraint to apologise for but the architecture that wins, once the loops are tight enough. Sovereign is where we find out.
Machines of loving grace — that serve your interior without ever being able to sell it. — The standing aim
The system's first incarnation: a multi-agent research-paper author.
Before it became a life-strategist, Sovereign was eight specialised agents authoring political-science papers — spec-author, source-curator, note-taker, lit-reviewer, debate-mapper, prose-author, methodologist, adversary — scored against a 2,401-paper golden set. It taught the lab the discipline the current system inherits: bounded faculties, grounded claims, a signed trace for every decision. The architecture is otherwise distinct. Preserved unedited.
A local-first intelligence substrate. The predecessor.
Eighty-seven engines on SQLite. An event-sourced spine. A ten-general strategy engine on the quaternion sphere. A measurement protocol that could test seventeen of its own hypotheses and refute eight of them. PLATO taught the lab what a bounded reasoning kernel actually needs to be useful, and where the limits of a human-life-as-dataset are.