Confrontation: The Citadel Stack's Threat Model against the Regulatory Threat Model
The Operator's Claim
The Citadel Stack — SYS-003 and its map MAP-003 — is built to prevent a capable system from escaping or subverting its binding. Every layer answers a loss-of-control question: can the system get out, can it negotiate, can it survive termination.
The Literature's Position
The regulatory and evaluation literature governs a different set of harms, and it is explicit about which:
- LIT-009 defines frontier risk around dangerous capabilities — the model's ability to do harm in the hands of a user, not its ability to escape its operator.
- LIT-012 (EU AI Act Art. 55) imposes duties keyed to systemic risk as defined in Art. 3(65): high-impact capabilities with foreseeable negative effects on public health, safety, public security, fundamental rights or society, propagatable at scale. The obligations are adversarial testing and mitigation of those risks — not containment of the model against its owner.
- LIT-008 identifies the incentive that arises once evaluation has consequences: the evaluated party benefits from understatement. That is an organisational threat, not a model-psychology one.
Where They Agree
Both treat capability as the variable that determines the required rigour, and both accept that a sufficiently capable system cannot be governed by asking it nicely.
Where They Diverge
The axes are different, and the Stack governs one of them. The literature's primary axis is misuse and uplift: what a bad actor can do with the system. The Stack's primary axis is loss of control: what the system does to its operator. These require different mechanisms. Containment does not prevent a legitimate operator from using a contained system to synthesise a pathogen; evaluation and access control do not prevent a system from subverting its monitor.
Three whole threat classes are absent from the Stack:
- Misuse and uplift. No layer in MAP-003 limits what the custodian may ask the node to do. The Stack is, by construction, a very secure way to give one operator unrestricted capability.
- Model-weight theft and insider threat as an organisational problem. CON-004 treats abliteration as a technical property of open weights. The literature treats it as an institutional failure: who has access, under what audit, with what liability. INS-007 already says weights in the wild is the dominant threat model; the Stack does not answer it, it assumes custody.
- The operator. INS-004 says a symbiosis posture is only credible if the humans are credibly bound, and INS-009 says necessity and convenience cannot be separated at enforcement time. Both are in the vault already, and neither is in the Stack's layer list. Every layer assumes the human side is trustworthy.
What Changes in the Vault
MAP-003 gains an axis note: the Stack is a loss-of-control instrument, and it should stop being described as a safety architecture without that qualifier. The misuse axis is not a refinement of the Stack — it is a second Stack, and the proposal has never said so.
CNF-007 takes the institutional half of this; CNF-009 takes the operator.