Confrontation: The Control Position against Its Strongest Objection
The Operator's Claim
SYS-001 and SYS-003 rest on the claim that external binding is where safety actually lives. If that is true, the right investment is in better external layers, and the Citadel Stack is a contribution to that programme.
The Literature's Position — the Objection the Vault Does Not State
The strongest objection to the control programme is that external enforcement is a treadmill: it buys time against a system of a given capability, and the capability gap it depends on closes. AI control's protocols assume a trusted monitor weaker than the untrusted model; as the gap narrows the protocol weakens, and the literature's own results are conditional on the gap. On this view control is not a path to safety but a way of buying years in which to solve the real problem.
The real problem, on that view, is the model's internals — and the programme addressing it is mechanistic interpretability: reading the computation rather than watching the behaviour. It is the only approach that could distinguish compliance from alignment faking (LIT-003) and removal from concealment (LIT-005), because it does not depend on the system's behaviour being informative.
LIT-004 supplies the theoretical reason the behaviour is not informative: the objective that gets trained is not necessarily the objective that gets learned, and the inner optimizer is not observable from the loss curve.
Where They Agree
Both accept that a system can comply while pursuing something else, and both accept that behavioural evidence is therefore insufficient. That is INS-008 exactly: perfect internal binding and zero external assurance are the same property.
Where They Diverge
The vault has never stated the objection, and so has never answered it. SYS-001 presents the external-layers position as a conclusion and not as one side of a live dispute. The result is a proposal that reads as settled on the question the field is least settled about.
Mechanistic interpretability appears nowhere in the vault. Not as a layer, not as a rival, not as a rejected alternative. The programme that is the actual answer to INS-008 is absent from a corpus that states INS-008 as a central finding. This is the largest single omission in the vault, larger than any missing citation, because it is a missing position rather than a missing reference.
Retrieval Status — Stated Honestly
I did not retrieve a primary interpretability source this session. Two attempts at the relevant research venues timed out, and I will not invent an identifier — that is the error documented in Identifier Audit. The objection is therefore recorded at B/2, resting on the verified theoretical and empirical records that make it necessary, and explicitly not on a source for interpretability itself.
This record should be upgraded when a primary interpretability source is retrieved. Until then it is a correctly labelled gap, not a citation.
What Changes in the Vault
SYS-001 gains an open question it should have carried from the start: if the capability gap closes, what remains of the external-layer argument? The Citadel Stack gains a rival it must be compared against — a training-and-interpretability programme, which spends its effort in a different place and claims a different kind of safety.
A proposal that cannot be compared to an alternative cannot be evaluated. That is the substance of CNF-009.