The Key to a Stable Governance Layer: Solve Foundational Social-Science Problems First :  Not the Other Way Around

AI governance does not start with governance.

In my view, this is one of the most common mistakes in the current conversation.

Too many efforts begin with compliance, audit trails, human review, dashboards, and controls. But if the underlying problem has not first been formulated coherently, all of that remains outer structure.

A stable governance layer is not a starting point. It is the result of understanding.

For those interested in the fuller argument, here it is.

In recent years, the discussion around AI governance has expanded dramatically: regulation, policy, risk management, kill switches, documentation, accountability, transparency. All of these matter. But a basic error often appears very early in the process: people begin building the governance layer before they have properly defined the problem that governance is supposed to govern.

That, in my view, is a sequencing error.

My approach is the reverse. Instead of starting from the outer layer: procedures, controls, standards, dashboards, and release gates: I begin by trying to formulate the underlying problem coherently at the level of social science. In other words: what is the structure of the decision process, who are the actual players, what are the relevant strategies, what kinds of failure emerge, under what conditions do those failures become stable, and how do institutions, language, incentives, and interpretation interact inside real decision-making?

Only after that does a governance layer become genuinely possible.

Without that prior step, what often emerges is only an illusion of control: there are metrics, but it is unclear what they really measure; there are procedures, but no clear logic of decision; there is human-in-the-loop, but it functions as a substitute for understanding rather than as a well-defined component inside a decision architecture.

Put simply: the system looks more governed than it is actually understood.

This is where the original framework I developed comes in: CEP :  the Central Equilibrium Problem. At this stage, I am not disclosing the full formalization, the internal terminology, or the complete analytical architecture of the framework, because it is still proprietary research material under development. Even so, the methodological principle I am presenting here is clear: a stable governance layer is not built from the outer shell inward; it is derived from a coherent formulation of the foundational problem in the decision process itself.

The central contribution of this framework is analytical discipline.

It forces a distinction between players, strategies, combinations, failure types, metrics, institutional rhetoric, and modes of enforcement. When those distinctions are not made, governance tends to become impressive on the outside but structurally weak on the inside.

This matters because failure in a system is not always an accidental malfunction. Sometimes it is the result of a stable structure. A system may persist in a problematic pattern not because no one noticed the problem, but because players, incentives, institutional language, and organizational logic keep converging on the same equilibrium.

If you do not begin there, the entire governance discussion remains at the level of patchwork.

In the work done here on the basis of CEP, the governance layer was not added from the outside. It was derived from within. That distinction matters.

Instead of asking first, “How do we supervise the system?”, the earlier question was: “What kind of decision process is actually taking place here, what can go wrong inside it, and how do we distinguish between a core failure and a surface-level symptom?”

Once that question is taken seriously, it becomes possible to build a more stable governance layer: not as organizational decoration, but as a mechanism with internal justification.

For example, it becomes possible to distinguish between Core and Shell: between problems that affect the logic of decision itself and those that are secondary manifestations or symptoms. It becomes possible to formulate metrics such as NFCI, not as another managerial KPI, but as an attempt to detect when language starts replacing explanation. And it becomes possible to define decision gates such as SHIP / RESTRICT / HOLD / ROLLBACK, not as generic operational preferences, but as direct extensions of structural risk analysis.

That is the difference between a system that has controls and a system that understands what it is stopping, why it is stopping it, and under what conditions.

This is also why the first step belongs to social science, not only to engineering.

The challenge of AI governance is not merely computational, and not merely regulatory. It touches foundational social-science questions: how actors make decisions under uncertainty; how institutions stabilize patterns; how incentives generate equilibrium; how rhetoric sometimes replaces explanation; and how a system can appear responsible and transparent while lacking any consistent ability to distinguish between risk, failure, justification, and decision.

Anyone who skips this level is likely to end up in one of three familiar outcomes.

First, compliance theater: everything is documented, but nothing is truly explained. Second, metric illusion: there are metrics, but they are not connected to a real decision mechanism. Third, escalation by fog: whenever things become difficult, the case is handed to a human: not as a principled step, but because the system lacks a coherent basis for judgment.

That is not governance. That is fog with procedures.

This is why I believe the sequence must be reversed from how it is often handled in practice: not to begin with the governance layer and only later try to fill it with substance, but to begin with a coherent formulation of the foundational problem: and only then derive the governance layer from it.

My claim is not that everyone working on AI governance must adopt CEP specifically. Nor is it that every governance discussion must become philosophical. The claim is narrower than that, but also more fundamental:

A stable governance layer requires, before anything else, a coherent formulation of the problem it is meant to govern.

Without that, what usually emerges is a system that looks orderly without being truly governed; sounds responsible without being able to explain itself consistently; and applies controls without possessing a clear logic of decision.

In my own case, the development of CEP led to a simple conclusion:

Stable governance is not the starting point. It is the outcome of a systematic understanding of decision structure.

First formulate the problem clearly. Then understand the players, the strategies, the failure modes, and the equilibrium dynamics. Only then build controls, gates, metrics, and release decisions.

Any attempt to shortcut that path may look efficient in the short term. In practice, it produces a governance layer with more forms than understanding.

Stable governance is not a substitute for understanding. It is the product of understanding.