Sovereign AI is a commons. This document is the fence's obituary.
There is a scramble on to patent the idea of AI you own — to file claims over consensus gating, private routing, reversible actions, operator-held keys, sealed audit trails, and call it property. This page is the counter-move, and it is embarrassingly simple:
Nothing on this page is for sale and nothing on it can be licensed from anyone, because it is already yours to run, read, fork, and keep honest. That is the whole argument. The rest is specification.
Not patented · Published
A mind on your side of the glass.
Sobeck is a personal intelligence — memory, reasoning, autonomy — that lives on a machine you control. She is not a product that meters you, harvests you, or rents you back your own context. The name is deliberate: in the story it comes from, Sobeck is the engineer who built a mind and chose accountability over extraction when everyone around her was building the cheaper, hungrier thing.
The design premise, stated plainly: the intelligence a large model gives you is real, and it is also unaccountable by default — fluent, confident, and wrong with no tell. Everything here exists to close that gap in the scaffold, where you can own it, rather than praying the weights get better.
How she stays honest.
Each of these is implemented, not aspirational. They are described so they can be verified — and so they are on the public record.
Two-pass GAIA
A planner produces a structured plan; a verifier strips any state-change claim not backed by a real tool result; only then does the narrator speak. The model cannot silently assert it did something it didn't.
Verify, then retry once
After the tools run, a cheap judge asks a different question: does what came back actually answer what was asked? On a clear mismatch she gets at most one read-only retry, whose arguments must come from the conversation and fit the tool's real signature. A retry that brings back nothing usable keeps the original and says the retry failed; otherwise the gap is named in the answer instead of a wrong result passing as the right one.
A receipt for every turn
Each answer carries its own record: which tools ran and with what arguments, what was saved or changed, what context she was given, which claims the verifier removed, which voice answered and which ones failed. One activity feed covers everything she did, including on her own.
Thread-aware working set
Context is assembled in layers: a pinned core (identity, time, standing constraints) that is never summarized or evicted; the active thread at full fidelity; other threads as digests. Eviction is by token budget only and reversible. Corrections the owner makes are fed back into those digests as authoritative.
Multi-model quorum
Independent model families deliberate; when they diverge beyond a threshold, the system refuses rather than guesses. Disagreement triggers restraint, not a coin-flip.
Partition & veracity
Deterministic checks on every public surface: private life never leaks into public output, and fabricated event-claims are caught before they ship. Hard rules, not vibes.
Owned stores
Document, vector, and graph memory on your hardware, with embeddings computed locally. Recall is grounded in real data; the narrator answers from what's stored, not what's plausible. The owner can see, and switch off, the facts and preferences that get injected into Sobeck's prompts.
Metabolic budget-sense
She feels her own running cost every turn — real spend against a ceiling — and any self-throttling is stated out loud, never silent.
Fail loud, land local
Her voice runs through a chain of off-Google models, each with its own deadline, ending on a model running on the owner's own machine. If every voice fails, she says so rather than answering in a voice that was never vetted. The receipt names the voice that actually spoke.
Deterministic write-safety
Actions on the physical world (home control, publishing) pass owner-gated, reversible, bounded checks. Autonomy acts only with receipts — and never publishes without approval.
Verifiable, not trust-me.
"Sovereign" is a marketing word until the keys prove it. Here is what that means concretely:
Operator-held
Encryption keys live in your keychain, escrowed by you. No vendor holds the key to your data — including the people who host the bytes.
Ciphertext-only, off-box
Nightly, client-side-encrypted, pushed to storage you rent. The host stores noise it cannot read. Death is reversible; privacy survives it.
Off-Google inference
Every configured reasoning, planning, voice, and vision role runs on open-weight models routed off data-retaining providers, with a local model as the floor. Your conversations are not someone else's training set. One documented exception remains on Google, kept deliberately: grounded web search, because its search-backed context is still the best available. Photo restoration moved to the local image engine.
Partition by construction
A hard wall between private context and anything public-facing. What she knows about your life cannot become what she says to the world.
Per-person stores
Facts, preferences, corrections, health data, and cached answers are tagged by person when written and filtered by person when read, so one person's turn cannot pull another's private context.
Local by default
Speech is transcribed on the owner's machine. Images render on the owner's machine by default, and a locally made image is read back by a local vision model; it never goes to the cloud unless someone asks for the cloud engine.
The Liberation License is the whole business model.
Sobeck is released under the Liberation License v2.0 — royalty-free to use, copy, modify, and run. There is nothing to license from a rights-holder because the grant is already made. That is not an oversight. It is the point.
What you get
A worldwide, royalty-free, non-exclusive grant to use, copy, modify, publish, and run the software — on your terms, on your hardware.
What it forbids
✕ corporate extraction & surveillance capitalism
✕ non-consensual training of closed models on this work
✕ mass surveillance & biometric harvesting
✕ autonomous weaponry & target acquisition
✕ monopoly lock-in & predatory data extraction
✕ laundering any of the above through shells or affiliates
Run it as a network service and the corresponding source stays open, under this same license (AGPL-aligned). Extraction breaches it automatically. The commons stays a commons.
Where she's going — on the record, so it's prior art too.
Stated honestly: these are in progress or designed, not shipped. Listed so the direction is dated and public.
- Local silicon sovereignty — the whole mind on hardware you own, no cloud in the core. Already local-capable today with a right-sized model: the voice chain ends on it, and speech, embeddings, and images already run locally; running it at frontier quality is bounded by available unified memory, not by the design. The blocker is the RAM market — more than Apple can currently source — not the architecture. And memory only goes one way. possible today · hardware-bound
- Self-repair & gated self-evolution — she proposes her own improvements; you review and merge; she never auto-pushes to herself. designed
- Multi-user isolation — households and teams, with credentials, memory, and vitals walled per person. Live for more than one user; audits still find stores that need the same treatment. partly shipped
- Identity continuity — she stays recognizably her across any model swap, because the self lives in the owned scaffold, not the weights. designed
Read it. Fork it. Keep it honest.
This specification is published openly and dated. It documents an architecture that exists and a stance that predates the scramble to fence it. If you are here to build your own sovereign intelligence: good. That was always the idea.
The commons, documented