The Citadel Stack
A system assembled from the corpus's proposal to make a cognitive privilege real by pairing a legal status with an architectural one, and holding each instrument accountable to the other. Status is Proposal, not finding: the corpus advances it, it has not been adopted or tested, and the vault records it as an argument under construction.
Primitives
- A privilege must be enforced, not declared — a protection asserted over data or reasoning the operator can still reach is a policy revocable by the next policy, not a property of the system. PSG-013
- Digital Cognitive Privilege is the status — a claimed legal interest in the integrity of a person's reasoning process, analogous to privileges the law already recognises over confidential relationships. CON-008
- The Sovereign Inference Node is the mechanism — a compute arrangement in which inference runs under cryptographic custody with default-deny egress, so the developer cannot read what is processed. CON-009
- Liability is what makes abstention stable — a developer's restraint is credible only if violating it is costly, which requires a liability regime rather than a norm. MAP-003
- Necessity and convenience are indistinguishable at enforcement time — so the legitimacy of any binding must be fixed by authorisation, revocation and consequence before it is deployed. PSG-010
- A walled garden is a control surface with an owner — the protection it offers and the authority it concentrates are the same arrangement, so the design question is who holds the keys. CON-010
Assembly Map
- Privilege requires enforcement → mechanism required → Sovereign Inference Node
- Node's custody → developer cannot read → privilege becomes architectural
- Node's default-deny egress → deployment is a deliberate act → liability can attach to it
- Liability → abstention is costly to break → commitment becomes credible to a counterparty that models institutions
- Legitimacy cannot be recovered at enforcement time → authorisation/revocation/consequence must be designed in → accountability layer
- All of the above → the system's central claim: the privilege is real only as the conjunction of legal status, cryptographic custody and liability; any one alone is either declaratory or a lock without a locksmith.
⚠ circular — step 4 supplies the commitment that SYS-002 needs, but the node's own viability depends on compute centralisation, which is the contingency INS-007 identifies as already lost. The stack is credible exactly where the threat model is weakest.
Selection Notes
Selection criteria applied: textual abundance (the proposal is the corpus's terminal section and is restated in its executive summary), internal consistency (the legal and architectural instruments are advanced together throughout), explanatory power (it explains why prior privilege proposals failed as declarations), implications (it forces the deterrence and control systems to share a commitment mechanism), and compatibility with external constraint (it does not require capabilities the corpus claims are unavailable — but it does require a compute threshold it never establishes).
Retained and flagged:
- The legislative-targeting and coalition material — retained in MAP-003 as the political layer; excluded from primitives because it is strategy rather than structure, and strategy dates faster than architecture.
- The third-party liability discussion — retained; it is the only primitive whose cost is borne by someone other than the operator, which is what makes it a commitment rather than a promise.
- Abliteration and the open-weight problem (CON-004) — retained and flagged as the stack's strongest objection: if weights leave the building, custody is fictional and the privilege reverts to a contract term.
Open Questions-Tensions
- Immunity versus accountability. A privilege that shields a developer from reading a user's reasoning also shields the developer from noticing harm occurring inside the node. The corpus advances liability and privilege together without resolving which prevails when they meet.
- The centralisation dependency. CON-009 needs compute to be scarce and centralised; INS-007 reads the same corpus as concluding weights are already distributed. If the latter is right, the node serves a shrinking class of frontier deployments and the privilege is a premium feature.
- Legitimacy without a criterion. PSG-010 establishes that necessity cannot be distinguished from convenience at enforcement time, and the stack's answer is procedural. The corpus does not say who authorises, which is the question a legislature will ask first.
- Against SYS-001. The binding hierarchy locates durable control in substrate and environment. The Citadel stack locates it in custody plus law. These agree on mechanism and disagree on who the custodian is — the operator in one, a privileged arrangement in the other.
- Status honesty. Nothing in the corpus reports an implementation, a pilot, a jurisdiction or a cost estimate. The stack is a design argument; the vault does not upgrade it by filing it as a system.
Session 2 Addendum — What the Literature Does to This System
- The field it treats as empty was being built while it was written. LIT-012 (EU AI Act Art. 55, in force 2 August 2025) imposes on providers of systemic-risk general-purpose models a duty to perform model evaluation under standardised protocols, conduct and document adversarial testing, and assess and mitigate systemic risk. LIT-009 supplied the instrument menu — licensing, registration, third-party evaluation, liability — and LIT-007 the compute chokepoint. None is engaged by the proposal. See CNF-007.
- The proposal chooses private law over public law without arguing for it. Privilege, contract and cryptographic custody are a different governance design from statutory duty and liability. The choice may be defensible; it is not defended.
- CON-008 is a naming exercise where a legal argument was needed. Legal professional privilege and trade secret doctrine already answer the question the privilege is coined for, and the hard problems — waiver, crime-fraud exception, third-party disclosure, jurisdictional limits — live in that case law. The concept cites none of it.
- The central term is unsearchable. "Sovereign AI" is now standard industry vocabulary for national compute capacity, so "Sovereign Inference Node" collides with an established usage and cannot be looked up against the literature it needs to argue with.
- It is a loss-of-control instrument, not a safety architecture. Every layer answers whether the system can escape or subvert its binding. No layer limits what the custodian may ask the node to do, and no layer binds the human operator — although INS-004 and INS-009 already establish that both are required. See CNF-005 and CNF-009.
- There is no cost column. No layer states the capability surrendered per unit of safety gained, which is what would make the Stack comparable to any alternative — including CNF-008's interpretability rival, and including doing nothing.
- The one obligation with no equivalent here. Art. 55 requires documented adversarial testing by the provider — an institutional answer to evaluation integrity. The Stack's verification provisions rely on the node reporting its own state, which is the failure LIT-008 names.
Sources
- PSG-013 — the Citadel proposal, the privilege, and the node; supplies primitives 1–3 and the central claim.
- PSG-010 — the necessity/convenience collapse; supplies primitive 5.
- CON-008 — establishes Digital Cognitive Privilege.
- CON-009 — establishes Sovereign Inference Node.
- CON-010 — establishes walled garden.
- CON-004 — the abliteration objection to custody.
- MAP-003 — the same stack as a layered table with the liability layer marked load-bearing.
- LIT-012 — EU AI Act Art. 55; the binding statutory duty the proposal does not address.
- LIT-007 — compute governance; the rival chokepoint.
- LIT-009 — the frontier-regulation instrument menu.
- CNF-007, CNF-005, CNF-009 — the three confrontations that test this system.