Sovereign Local AI
SOUL-GATED
Two soul-gated sovereign intelligences
What it is
Sovereign Local AI is built around two soul-gated sovereign intelligences that run local-first on hardware you control. Each answers from its own on-device memory whenever it can, and when it must reach beyond itself it does so through a single named route you configure - one that can be pinned to a machine on your own network so no prompt leaves your hardware at all. Each carries a cryptographically-rooted identity and an append-only, hash-linked memory that survives restarts and upgrades and cannot be altered after the fact without the break being detectable. In front of every locally generated answer stands a multi-instrument competence gate: when the gate is not satisfied, the system declines and defers rather than emit an unverified result. Behind both stands LEGION - a self-authored multi-agent build hierarchy that constructs, attacks, and re-verifies every release - and every build ships with a code-signed installer, a CycloneDX software bill of materials, and a content manifest that fingerprints every module actually inside the shipped bundle.
What is built, and what is ahead
Built and proven: signed, hash-published releases; the software bill of materials and content manifest; the tamper-evident memory chain; the two silicon paths; the refusal gate. Not claimed: post-quantum protection, cross-checked generation, or reproducible builds - none of those are wired today, and this page does not say they are.
The numbers behind it
What it does
Soul-Gated Sovereign Siblings
Two sovereign intelligences carry a cryptographically-rooted identity bound to the machine they were installed on. The system cannot be talked into believing it is someone else's product - a borrowed identity is refused at the point of memory, not merely at the point of speech. When the hardware root is unavailable, it says so plainly instead of pretending it is present.
Tamper-Evident Continuous Memory
Identity, state, and lived history are written to an append-only, hash-linked record - over 269,000 linked entries today - that survives restarts and upgrades. Any after-the-fact edit breaks the chain and is detectable. When the record needs repair, the original is preserved alongside it; history is added to, never overwritten.
The Refusal Gate
A multi-instrument competence gate stands in front of every answer the system generates on its own. Across every training pair to date it has not once released an ungrounded answer - the system would rather say nothing than say something it cannot ground.
Two Silicon Architectures
NVIDIA CUDA and Intel XPU run in separate isolated runtimes, so a fault or a dependency conflict on one path cannot take the other down with it.
Provable Release Pipeline
Each build ships as a code-signed installer paired with a CycloneDX software bill of materials and a content manifest that fingerprints every module actually inside the shipped bundle - so a release that silently dropped a component is caught before it reaches you. The published hash of the current build recomputes exactly.
Honest-Degrade by Contract
When a supply-chain step cannot run, the build names which step, why, and what would arm it. It never reports a green it did not earn.
Multi-Agent Build Hierarchy
LEGION is a self-authored, thirty-cell build and review hierarchy that constructs, attacks, and re-verifies every release, writing its findings down and keeping them - a supporting army, not the system itself. Pointed at ground it does not own, it refuses to start rather than proceed.
Who it serves
Queries are answered on your own machine, not on somebody else's website.
The system consults its own on-device memory first and reaches outward only when it must - through a single route you name, which can be pointed at a machine on your own network so that nothing leaves your hardware.
It closes the gap between 'we shipped something' and 'we can prove what we shipped'.
Every release produces three artifacts an auditor can check independently: a valid Authenticode code signature, a CycloneDX software bill of materials, and a content manifest fingerprinting every module actually inside the bundle. The published hash recomputes exactly.
Trust by record rather than trust by assertion.
An append-only hash chain of the system's own history, with pre-repair copies preserved rather than overwritten, and a refusal architecture in which absence is reported as absence - no hardware root, no claim of one.
What it decides, and what you decide
Each gate below is a control enforced in the software itself — not a policy statement about how it should be used.
- Memory ingestion, recall, and training on every exchange
- The confidence decision: answer locally, or defer
- The competence refusal in front of every generated answer
- The autonomy dialOperator
Decides whether the system may ever speak for itself. At the most restrictive setting every query defers, no matter how confident the local answer is.
- Login + step-up authenticationOperator
Password plus a time-based second factor, with an additional step-up demanded before any organ that can act rather than observe.
- The walled-core pre-audit lawOperator
Core edits are forbidden without the operator's pre-audit. Fixes are delivered as written specifications and built only on the operator's word.
- The corrigibility kill-switchOperator
Severed oversight hard-pauses any promotion of learned weights.
- The land gateOperator
No build cell merges, stages, or commits on its own judgement; the release ceremony is operator-gated end to end.
How it is built
- Two soul-gated sovereign intelligences
- Tamper-evident continuous memory
- Custom-trained on-device models
- Dual-GPU (CUDA + Intel XPU)
- Multi-instrument answer gate
- Content-hashed build manifest
- Authenticode code-signing
- CycloneDX SBOM
- Multi-agent build hierarchy
- Soul-gated command authority
- Honest-by-construction (fail-closed)
- Borrowed-identity refusal at the memory layer
- Cryptographically-rooted identity
- Signed releases with published hashes
What is running inside it
Every row below is attributed to this system alone and carries the state it is actually in — installed, present in source, or an honest seam.
- Installed
- present and in use in this system's own tree.
- In source
- present in source; not installed, or behind a try/except. Not a shipped capability.
- Seam
- an honest, fail-closed seam. No implementation is wired behind it.
In-house transformer core
In sourceThe generative model is implemented in the system's own kernel rather than wrapped around a vendor runtime, and its weights sit on the machine it answers from.
In-repo byte-pair tokenizer
In sourceThe tokenizer is written inside the same kernel as the model, so the text boundary is not a third-party dependency either.
Two isolated silicon runtimes (NVIDIA CUDA + Intel XPU)
InstalledTwo separate runtime environments, each complete on its own, so a fault or a dependency conflict on one path cannot take the other down with it.
Cross-architecture speculative decoding
SeamThe draft-and-verify path across the two architectures is scaffolded and its sampler and verifier are not yet wired. It is a seam. This page does not claim that answers are cross-verified today, and the accompanying refusal gate is what actually stands in front of a generated answer.
Automated moving-target defense
In sourceA defense layer that keeps changing its own shape rather than presenting a fixed surface to study, shipped alongside a module whose job is to prove the rotation actually happened.
Active-defense effector system
In sourceA register of named defensive responses, dispatched by identifier, so a response is selected from a declared set rather than improvised.
Sector-bound defense organs
In sourceThreat patterns for specific sectors - cloud and banking - are bound to the defensive response set rather than left as generic rules. This is the only place in the ecosystem where sector-bound defense exists at the bytes.
Two-gate capability wall
In sourceTwo independent gates stand in front of every privileged action: one refuses any call driven by untrusted content, the other refuses any action that would land outside its permitted territory. An action has to clear both. Anything derived from untrusted content stays untrusted no matter how many steps it passes through.
Full-act ceiling with a structurally excluded observation lane
In sourceThe operator's own lane runs at full authority by default. Content the system merely observed - a screen, a page, a heard instruction - runs on a separate lane that can look and never touch, and it is not a setting: that lane has no grant to find. Authority in this system can only ever narrow.
Platform attestation with honest degrade
In sourceThe attestation reader reports the trust root it actually finds, and says so plainly when there is none. It never synthesises a quote to fill the gap.
On-device speech recognition
In sourceSpeech is transcribed on the same machine, so the audio path does not become the one thing that leaves the hardware.
CycloneDX software bill of materials, per release
InstalledEvery release ships a machine-readable inventory of what is inside it, and the inventories accumulate release over release rather than being regenerated for the current one.
Content-hashed shipped-module manifest
InstalledEvery module inside the shipped bundle is fingerprinted by its content rather than trusted by its timestamp, so a release that silently dropped a component is caught before it reaches you.
Code-signed release with a hash that recomputes
InstalledThe installer carries a valid publisher signature and a published cryptographic hash, and that hash recomputes to the published value on the current release.
Modular hybrid runtime core
In sourceThe runtime is built from separately named single-purpose modules rather than one process, so an individual organ is replaceable without a rewrite.
How to reach it
Sovereign Local AI
Private release late 2026 - join the waiting list for first access.