The Locus That Was Designed Away: When Delegation Is Engineered to Make Accountability Unassignable

The third application of the completed framework and what it reveals about the precondition the framework has always quietly depended on

A decision is made. Something goes wrong. The natural question (who decided) finds five honest answers, all the same: “I only followed the previous step.” Nobody is lying. And nobody decided.

This essay treats the structure that produces that result. Not by accident. By design.

The prior edition examined the unmapped chain; delegation that leaks because the institution lost sight of where accountability for its data descends. That was a problem of absence. Something that should have been governed was not. The response was to map what had been left dark. The chain becomes governable once the institution can see it.

There is a more serious problem, and it is not an absence. It is an architecture. Some delegation chains are constructed so that, at the end, no single actor can be identified as having decided. Each link receives an output from the previous one, acts on it within the narrow scope assigned to it, and passes its result to the next. Each link can say truthfully that it exercised no independent judgment, because none was required of it. The chain runs from input to outcome. A decision is produced. Trace it back to find who made it, and the chain dissolves in your hands. Every link points to another. None holds.

This is not the unmapped chain rendered worse. It is a different structure. The unmapped chain can be mapped, and once mapped, governed. The diluted chain offers nothing to map. It is fully visible and still ungovernable, because it was engineered to have no point at which authority concentrates. You can see every link. You still cannot find the decision.

Naming what the problem destroys

The conceptual move this essay introduces is not to name the problem. It is to name what the problem destroys.

Call it the locus of attributable authority. The locus is the point, inside any decision structure, where responsibility for a choice can be unambiguously assigned to an identifiable agent. It is the silent precondition of nearly everything that ordinary language calls governance. Without locus, there is no one to approve, refuse, veto, or be called to account. There are decisions. There are no deciders.

The landing point states this in a sentence: authority that cannot be located cannot be held, and cannot be interrupted.

The reason the locus matters more than it appears to is that it is not, in normal conditions, an explicit element of governance design. Healthy organizations have locus the way healthy bodies have a circulatory system — present, functioning, unmentioned. The locus becomes visible only when it is destroyed, and when its destruction prevents the act of holding to account that everyone assumed would be possible. By the time the locus is missed, the structure that destroyed it is already in place, and the people who built it can each truthfully say they were only following the previous step.

Diluted delegation chains attack the locus directly. Not always maliciously. Often through inertia, mutual protection, or aversion to the weight of deciding. Dispersing responsibility is the path of least resistance when no one wants to hold it. The structure that results protects all participants from the burden of having decided. And, by doing so, it ensures that no one is actually governing.

One distinction has to be drawn here, because it bears on the coherence of everything published over the past months. The locus of attributable authority is not a fourth coordinate of the framework. The framework remains complete in three: Systemic Augmented Intelligence, Architectural Legitimacy, Institutional Dialogue. What this essay introduces is the recognition of a precondition the framework has always quietly required. Architectural Legitimacy presupposes a point where authority can be exercised. The Architecture of Refusal, formalized in the February 2026 working paper, presupposes a point where refusal can be issued. Cognitive signature presupposes a sujet who can sign. All three rest on locus. Naming the locus is not adding a coordinate. It is making explicit what the existing coordinates rest on.

A phenomenon older than the technology that intensifies it

Engineered dilution is not new with AI. It is older than the field by many decades, and recognizing this is essential, because treating it as a property of automated systems alone underestimates how naturalized the structure has already become.

Regulatory architectures deliberately fragmented across multiple agencies with overlapping jurisdiction, conglomerates organized in holding-company layers where authority dissolves at each ascending level, compliance regimes distributed across so many committees that no single committee can be held accountable for the system as a whole; all of these are forms of engineered dilution that predate the first useful AI model by half a century or more. The pattern is recognizable across domains. Whenever responsibility for a consequential decision carries enough cost — financial, legal, political, reputational — that no single actor wants to bear it, organizations evolve structures that disperse the responsibility until it stops being assignable.

This is engineered dilution in its mature human form. It is sophisticated. It usually requires careful organizational design, sustained political effort to maintain, and explicit institutional knowledge among the participants that the design is meant to do what it does. It is also rare in any single organization, because it costs effort to build and effort to keep running.

What has changed is not the existence of the pattern. What has changed is how easily the pattern can now be reproduced and how often it appears not by deliberate design but as the default output of a particular kind of architecture.

Multi-agent AI as the technical form of mass-produced dilution

Multi-agent AI architectures have made trivial what once required careful organizational engineering. The basic structure of a multi-agent system is, in technical terms, a chain of algorithmic agents that pass outputs to one another and produce a final result by composition. Each agent performs a narrow function. Each takes the previous agent’s output as input. The system as a whole produces a decision (recommend, approve, reject, escalate, deny) and that decision emerges from the composition of agents none of which, individually, can be said to have decided.

Note that this is not, in itself, a defect of the technology. Multi-agent systems are powerful precisely because they decompose hard problems into manageable specialized components. The decomposition that makes them effective is the same decomposition that makes them, by default, instances of engineered dilution. The default is the dilution. Avoiding it requires deliberate architectural work that runs against the technical grain.

The consequence is that engineered dilution has migrated from being a sophisticated organizational achievement, undertaken with intent, to being the standard structural output of contemporary AI deployments. Institutions that would never have built a deliberately fragmented compliance regime are now adopting multi-agent systems whose composition produces, as a side effect, exactly the structural property a fragmented compliance regime would have aimed at: a decision flow with no locus.

This is the urgency behind the concept. The problem is not that someone is deliberately destroying locus. The problem is that locus is being destroyed as a byproduct, at scale, without anyone choosing the destruction. By the time the locus is missed, the systems that destroyed it will be embedded across the decision infrastructures of multiple sectors. The cost of restoring locus after the fact is structurally higher than the cost of preserving it at design time and the longer the delay, the higher the cost.

The framework response, and the completion of the arc

This edition closes a three-edition arc whose internal structure is worth making explicit, because it reveals what the framework has been doing.

The arc opened with HNDL (Essência AI Insights Edition #29), which established the temporal horizon. The question there was when institutional judgment becomes auditable: in the future, when encrypted archives become readable, the cognitive structure of past decisions becomes visible. The arc continued with the unmapped chain (Essência AI Insights Edition #30), which established the spatial horizon. The question there was where in the institutional structure substantive judgment must be present: at every link of the delegation chain through which accountability descends, including the links the institution has lost sight of.

This edition establishes the intentional horizon. The question is who — within any given decision structure — can be held accountable for the result. Under engineered dilution, the answer, by construction, is no one. The temporal horizon (#29) addresses when the failure of judgment becomes visible. The spatial horizon (#30) addresses where the failure occurs. The intentional horizon (Essência AI Insights #31) addresses whether there is any actor against whom the failure can be charged at all.

Three horizons. One framework. The triad — Systemic Augmented Intelligence as cognitive substrate, Architectural Legitimacy as the form of authority, Institutional Dialogue as the form of continuous relation — responds to all three without growing a new component. Each coordinate, examined under the pressure of engineered dilution, reveals the same dependency: it presupposes the locus.

Architectural Legitimacy presupposes a structure within which authority is exercised — and authority cannot be exercised by no one. SAI presupposes a substrate in which human judgment is documented at the cadence of execution — and judgment cannot be documented if no human has been positioned to exercise it. Institutional Dialogue presupposes parties who can sustain a continuous relation — and a structure designed to have no identifiable actor at the relevant nodes has no parties to sustain dialogue with. All three coordinates require locus. Engineered dilution attacks all three simultaneously by attacking the precondition all three share.

The boundary of where this applies

This applies to decision structures of consequence — structures where the outcome of a decision carries real institutional weight, financial, fiduciary, regulatory, civil, or criminal. It applies with particular force to multi-agent AI systems operating in such contexts: credit assessment, medical decisions mediated by AI, large-scale content moderation, critical infrastructure governed by agent composition.

It does not apply to low-consequence decisions where attributing authorship would cost more than the value protected. Invoking locus of attributable authority for every microdecision inside a routine system is to waste the concept. The concept earns its weight on decisions whose outcomes can be held against the institution that produced them.

The practical implication, and a horizon worth marking

The conclusion follows the pattern of the prior two editions, and gains its full weight only at the end of the arc. Knowing when judgment becomes auditable (the lesson of #29) requires knowing where in the structure judgment had to be present (the lesson of #30). Knowing both requires knowing who can be held accountable for the structure as a whole. Each horizon presupposes the next.

The practical implication is that designing a decision architecture — institutional or technical — now requires preserving the locus of attributable authority deliberately, against the default tendency of contemporary multi-agent systems to dissolve it. This is not extra work. It is the work that the framework has always been describing, made explicit by the conditions that now require it to be explicit.

Three horizons close this arc. The temporal exposure of institutional judgment, the spatial structure of delegation, the intentional architecture that determines whether accountability is even possible. Each is a face of the same underlying problem: the continuity of substantive judgment, embedded in institutions, across time, structure, and assignment.

That underlying problem has an analog at the individual level, and recognizing it is the right way to close the arc. The same conditions that erode institutional locus also operate on the individual professional whose use of AI produces artifacts that exceed the competence the professional has actually built. The structure is the same. The artifact exists; the subject who produced it has not accompanied the continuity the artifact requires. AI, in this particular sense, amplifies what one is and erodes what one tries to be. This is a horizon for another moment. It is named here to register that the problem this arc has traced through the institutional level has a counterpart at the individual level, and that the two are not separate phenomena. They are the same phenomenon at different scales.

The next edition opens a new arc — the mechanism by which legitimacy is conferred, and the structures of approval, verification, and validation that have become both more necessary and more difficult in a moment when the act of producing has stopped being a reliable signal of competence.

Essência AI Insights Governance, systemic risk, and institutional infrastructure for emerging technologies