Locally Compliant, Collectively Off ── Moving the unit of governance from the agent to the population

Passing Component by Component Is Not Passing

Financial institutions are beginning to run agentic workflows in credit, fraud, collections, compliance and operational control. Governance, for now, is applied component by component. Each model or agent is specified, tested, authorised and monitored on its own.

What this paper presses on is the point at which that shape runs out. Where institutional risk arises from the joint behaviour of components that are each individually acceptable, no amount of component-level approval adds up to an acceptable whole. The authors name the gap constitutional non-compositionality: local compliance checks need not compose into acceptable collective outcomes such as bounded disparate impact, market integrity, or traceable accountability.

The Proposal Moves the Unit of Governance

ARIA is presented both as a finance-specific reference architecture and as a falsifiable research agenda. Its core is a change in the unit of governance. Instead of examining agents one at a time, it treats the population of agents as the thing observed.

To handle a population, six capabilities are distributed across three planes: normative-accountability, execution-control, and assurance-learning. They comprise policy specification, population-level monitoring of observed versus expected behaviour, bounded authority, runtime containment, adaptive policy change, and preserved human oversight competence. The last of these is the one worth pausing on. What is listed as a capability to maintain is not an oversight mechanism but the competence of the people doing the overseeing — the thing that quietly erodes when a machine takes over the routine cases.

What the Abstract Claims

What is offered as evidence is illustrative rather than conclusive: two simulations. One shows thin-file exclusion arising through a shared signal while local controls are functioning as designed. The other shows that monitoring the observed distribution against the expected one raises a warning earlier, in a constructed drift regime^[1].

The authors map these controls onto the evidence needs of fair-lending supervision, the EU AI Act, model-risk management and conduct supervision^[4]. The paper then closes not with a claim of production effectiveness but with a validation agenda. That restraint is unusual and makes the work easier to report honestly: the authors themselves state that what they have shown is an arrangement, not evidence that the arrangement works.

The primary source is a preprint that has not been peer reviewed. Anything beyond what is set out here waits on the full text and on review.

How Local Compliance Comes Apart in Aggregate

The first illustration in the abstract has a shape that will be familiar to anyone in regulated review work.

No component is in breach. The institutional outcome is nonetheless unacceptable. This is why the authors argue for raising the level of monitoring from component to population, and for watching the gap between the observed distribution and the expected one rather than the correctness of individual decisions. It is a change of direction in what gets measured, not simply more measurement.

What Is Not New, and What Cannot Be Read

The fallacy of composition is not a new observation. Practices that are sound in isolation producing fragility when everyone leans on the same indicator or the same data is a pattern financial supervision has handled repeatedly. What is new here is the application of that pattern to populations of agents, and the attempt to translate it into an allocation of controls and into evidence a supervisor can ask for.

A great deal cannot be read from the abstract. How the simulations were constructed and at what scale, how the observed-versus-expected gap was normalised, and how much earlier "earlier warning" actually is — within the abstract, the conditions are not given. Reference architectures also carry an inherent weakness. Listing capabilities and being able to verify from outside that those capabilities were implemented are different things. That the authors close with a validation agenda instead of an effectiveness claim suggests they are aware of it.

Three Papers on the Same Question at the Same Time

This subject arrived as a cluster. Within the harvest window, a paper on a tiered multi-agent framework with distributed-ledger audit trails for regulated financial operations, and a paper on a framework for insurers written against both prudential regulation and the AI Act, appeared alongside it^[2][3]. Three in total.

What the three share is a starting point. None of them begins from whether to adopt agents; each begins from what evidence will remain after adoption. There are no reader votes and no following implementations here, so this is not a cluster formed by public attention. It shows researchers digging in the same place at the same time, and nothing more should be read into it. A cluster forming says nothing about whether the claims hold.

It Reads as Finance and Applies to Material Review

When AI is used in regulated work, the unit of authorisation usually matches the unit of tooling. A system is specified, tested, authorised, and recorded. Support for promotional material review will be handled the same way.

Bringing this paper's argument in adds a question. When everything leans on the same basis, which way does the population drift? If review-support agents all rely on the same reference documents, the same criteria and the same body of past cases, each may be defensible while the organisation's judgement as a whole tilts in one direction. Tilt towards severity and the people producing material are worn down; tilt towards leniency and deviations pass. Neither is visible in component-level testing.

In practice, then, what needs watching is not only whether individual determinations were right. It is the distribution of findings over time: which categories of finding grew, which shrank, and whether that happened because the material changed or because the supporting machinery started facing the same way. Keeping records that can tell those two apart is the first step towards supervising behaviour at the level of a population. On that point, the authors' decision to count human oversight competence as a capability to be maintained sits closer to practice than it first appears.