Confrontation: The Missing Operator against the Organisational Threat Model
The Operator's Claim
The Citadel Stack's layers bind the system: the model cannot escape, cannot negotiate, cannot survive termination. The humans in the arrangement appear as the party for whom the binding is constructed.
The Literature's Position
The literature treats the human side as the threat surface:
- LIT-008 locates the sandbagging incentive in the evaluated party, not the model — the party that benefits from the evaluation understating capability. That party is an organisation.
- LIT-009 builds its instrument set around firms: licensing, registration, reporting, third-party evaluation, liability. Every instrument exists because the developer is not assumed trustworthy.
- LIT-012 places the duty on the provider. The regulated entity is the organisation, and the obligation is documented and adversarial by design.
Where They Agree
The vault already contains the correct instinct and does not use it. INS-004 holds that a symbiosis posture is only credible if the humans are credibly bound. INS-009 holds that necessity and convenience cannot be separated at enforcement time, only at design time. Both are organisational insights sitting in a system that binds only the machine.
Where They Diverge
The realistic failures are human and the Stack does not model them.
- The human with root. Every layer in MAP-003 is enforced by someone who can disable it. The Stack has no answer to the custodian who is coerced, bribed, or simply persuaded — the failure PSG-010 touches obliquely and the layer list does not carry.
- Jurisdiction shopping. A custody regime that depends on one legal order is defeated by incorporating elsewhere. CON-009 assumes a node under a sovereign, and says nothing about which sovereign or why they would not compete for the business.
- Insider threat as the dominant vector. INS-007 already says weights in the wild is the dominant threat model. The Stack's answer is custody — but custody concentrates the weights in one place with one set of administrators, which is the classic insider-risk trade.
And there is no cost accounting at all. No layer states what capability is surrendered per unit of safety gained. Without that, the proposal cannot be compared to any alternative — including doing nothing, which has costs the proposal never names. A regime that costs a great deal of capability and buys an unmeasured amount of safety is not obviously better than a regime that costs less, and there is no page in the vault on which that comparison could be made.
What Changes in the Vault
SYS-003 gains a fourth open question, alongside the institutional one: who binds the operator, and what does each layer cost in capability? INS-004 and INS-009 are promoted from observations to unmet requirements of the Stack.
A cost column is the minimum addition that would let the proposal be evaluated rather than admired. It is also the thing that makes the comparison to CNF-008's rival programme possible: two designs, two costs, two claims about what safety they buy.