The Governance Problem That Rules Cannot Solve
Gillian K. Hadfield, Rakshit Trivedi, and Dylan Hadfield-Menell published a technical research agenda in March 2026 that starts from an observation most AI governance work quietly sidesteps: democracy is not a rulebook. It is a complex adaptive system, constituted by the concurrent adaptation of millions of independent agents responding to and anticipating each other's behavior within a shared normative order, a "dancing landscape," as they put it.
The implications for AI governance are significant. If democratic order is dynamic and adaptive rather than static and rule-governed, then any approach to AI alignment that encodes democratic values as fixed parameters will almost surely fail to protect democratic stability, not because the parameters are wrong, but because they are incomplete by definition. No rulebook can anticipate every circumstance. No contract can specify optimal behavior in every future state. And the norms and laws that fill those gaps are themselves incomplete, subject to ambiguity, interpretation, and evolution.
This is what the authors call the incompleteness problem, and it is not a bug in current governance approaches; it is a structural feature of how normative social order actually works. Human societies have addressed it over millennia by developing classification institutions: identifiable entities that serve as authoritative sources for resolving normative ambiguity, filling gaps, and settling conflicts, backed by patterns of third-party punishment. Courts, regulatory bodies, standards organizations, entities that produce legible, stable classifications of what is permitted, contested, and punishable, and that earn enforcement support precisely because they are structured as neutral, open processes rather than partisan ones.
AI agents operating at scale will face this incompleteness problem. They will need to make continuous decisions, about product design, employment, marketing, regulatory compliance, that are embedded in a normative social order they cannot fully pre-load. What the authors call normative competence is the capacity to read that normative environment at inference time — at the moment of action, not only before deployment — and adjust behavior accordingly, in effect by simulating an impartial spectator that predicts how appropriate institutions will classify behaviors and how likely punishment is to follow.
The paper's technical research agenda identifies what needs to be built: AI systems with normative competence, and new normative institutions, "model specification institutions," that operate at the speed and scale of AI, provide inference-time access to authoritative classifications, and are structured with the legal attributes required to earn enforcement support: stability, generality, clarity, neutrality, impersonal reasoning. These institutions must be independent of the AI developers whose systems they govern, and the authors are explicit that they do not yet exist.
Beyond Democratic Framing
The paper writes explicitly from a democratic governance framing: it takes as its premise that preserving democratic social order is the central value AI governance must protect. That framing is not universal. Safety-oriented work focuses on catastrophe avoidance and is compatible with both democratic and non-democratic political orders, while public-interest approaches prioritize societal welfare without tightly specifying democratic process. Yet the technical problem the authors identify, incompleteness, the need for inference-time governance, and the structural requirement for independent classification institutions, arises under any governance value framework. The architecture they sketch is developed in service of democracy, but it is better understood as a general response to a governance problem that every large-scale AI deployment faces.
The Inference-Time Gap
The major approaches to AI governance currently emphasized by frontier labs differ in method, but they share a common structural assumption: that governance can be adequately specified before deployment. Constitutional AI encodes values and behavioral constraints during training. Preference aggregation elicits human values and encodes them as reward signals. Rule-based frameworks specify prohibited and required behaviors in advance. Law-following AI trains systems to recognize and comply with legal rules.
Despite those methodological differences, they share the assumption that the relevant values, norms, and constraints can be identified, encoded, and applied in advance, and that checking behavior against those encodings is sufficient.
Hadfield's framework explains why this assumption fails. No training-time encoding can track the evolution of the normative environment after deployment begins. The incompleteness is not resolvable by making the specification more detailed, because the normative environment itself is dynamic, produced continuously by the behavior and beliefs of agents within it. You cannot fully specify in advance what will not be known until the system is operating.
The result is a structural gap across every dimension of deployment. A system governed by pre-deployment specification cannot ask, at the moment of action, what the current normative context requires. It cannot detect that a constraint has become contested, or that a new class of behavior has emerged that the original encoding did not anticipate. It is, in Hadfield's precise framing, a training-time solution to an inference-time problem.
This is not a critique of the intentions behind these approaches. They represent substantive efforts to govern powerful systems. But the limitation is not a design failure. It is epistemological: no specification can anticipate what does not yet exist. The normative environment is generated continuously by agents acting within it, which means the gap between pre-deployment encoding and deployment reality cannot be closed by specification alone.
This limitation can only be addressed by systems capable of normative inquiry at the moment of action: continuously reading context, assessing it, and responding accordingly. That is, after all, how human reasoning navigates the same problem.
This is where theory and infrastructure converge. Ludulluu was not built from this framework; the two lines of reasoning developed independently. Hadfield's framework proposes institutions that produce authoritative classifications through human deliberation. As a research agenda, it does not specify how competing behavioral preferences are translated into computable thresholds and applied at inference time at the speed of deployment. Ludulluu is building that infrastructure.
Inference-Time Governance as an Engineering Constraint
If Hadfield, Trivedi, and Hadfield-Menell are right about the incompleteness problem, then the space of viable governance architectures is narrower than it looks. Any deployment of powerful models will still need principles, guardrails, and training-time constraints. But those alone cannot govern what actually happens when systems operate inside a live, evolving normative order. The gap at inference time does not disappear because a lab adds more rules, more datasets, or more human feedback; it is structural to the problem they are trying to solve.
The governance challenge is to move from policy to infrastructure: from rule sets to the computational capacity to exercise normative agency at the speed of deployment. Normative agency requires infrastructure: systems that can execute, monitor, log, test, and audit governance in operation. This makes governance legible, verifiable, and scalable. It also creates the evidentiary basis for trust, without which technical capability alone cannot earn legitimacy, sustain adoption, or be governed at scale.