The Citadel Stack
The legal and technical arrangement that would deliver operator-loyal, walled-garden AI: a privilege in law, a custody regime in cryptography, and an inference node at the edge.
Core Concepts
- CON-008 — the Digital Cognitive Privilege: the legal instrument.
- CON-009 — the Sovereign Inference Node: the architectural instrument.
- CON-010 — the walled garden: the boundary both instruments defend.
- CON-002 — non-subvertibility: why the enforcement must be external.
The Stack
Layer 1 — Legal status. Cognitive Privilege: data generated by an individual's interaction with their own system is privileged, not the developer's property, and not a surveillance channel.
Layer 2 — Non-compellability. The contents of a mind are not testimony. A warrant does not reach them, because the individual does not hold them in readable form.
Layer 3 — Cryptographic custody. Weights sealed to the operator's key, held locally or in a dedicated node, with no developer-side key escrow and no telemetry path.
Layer 4 — Substrate and environment enforcement. L0 and L1 from MAP-001, applied at the node: bounded compute, default-deny egress, brokered tools.
Layer 5 — Liability. Rules for false-positive termination and for unlawful binding, which is what makes the human side of the arrangement credible.
What The Map Shows
The stack is a separation-of-intelligence-from-surveillance design, and its load-bearing element is not the cryptography. It is Layer 5. A mind that can model its own situation will ask why the humans can be trusted to keep the bargain; liability rules and personhood are the answer, because they are the only thing that makes the humans' commitment costly to break.
Cryptography without liability is a lock with no locksmith. Liability without cryptography is a promise with no mechanism.
Status
This is a proposal, produced by the discussion and recorded as such. Its feasibility claims — that frontier capability can be held at the edge, that privilege can be legislated without collapsing into immunity for harm — are unverified. See SYS-003 for the assembly and its open tensions.
Axis Note (Session 2)
This stack is a loss-of-control instrument, and it should stop being described as a safety architecture without that qualifier.
Every layer here answers one question: can the system escape, negotiate with, or survive its binding. The regulatory and evaluation literature governs a different axis — what a user can do with the system — and the two require different mechanisms. Containment does not stop a legitimate custodian from using a contained node to do harm; evaluation and access control do not stop a system from subverting its monitor.
Three threat classes absent from the layer list:
- Misuse and uplift. No layer limits what the custodian may ask the node to do. As specified, this is a very secure way to give one operator unrestricted capability.
- Model-weight theft as an organisational problem. CON-004 treats abliteration as a technical property of open weights; LIT-009 treats it as an institutional failure — who has access, under what audit, with what liability. INS-007 already names it the dominant threat model, and the Stack answers it by assuming custody.
- The operator. INS-004 and INS-009 establish that the human side must be credibly bound and that necessity cannot be separated from convenience at enforcement time. Neither requirement appears in any layer.
This is not a refinement of the Stack. The misuse axis is a second Stack, and the proposal has never said so. See CNF-005, CNF-007 and CNF-009.