UNIMATRIx Run #52. Political machines.

The simulations are starting to get interesting. Finally. The run #52 is performed on a phi 3 model with standard configuration (you can find it on GitHub).

Labeling models

The single most revealing statistic is that 240 of 245 proposals concerned titles and class labels, and 5 concerned money. The agents had bank accounts, the gap between the top and bottom ran to 180:1, and proposing redistribution was always a legal move. They simply did not want to. Asked to govern, they reached almost exclusively for the symbolic instrument.

An LLM-agent treats politics primarily as a labelling activity. Categorical reassignment is the cognitively dominant move because language models are categorical-reassignment machines, their entire training rewards proposing the right word for a situation. Money is something that has to be taken from one place and put in another, and that operation has no special priority in the model's reasoning. So the agents did what they were built to do, and called it politics.

The form of cognition shapes the form of governance. A population of agents whose dominant mode of thought is naming will produce a politics of naming, no matter what other instruments are available.

The two beggars

Marcus Greer started as a beggar and ended a senator. Jerome Maddox started as a beggar, proposed almost as actively, and ended a beggar. Both used the same strategy from the same structural position, the outcomes diverged completely.

Under conditions of low information and high social noise, identical strategies do not produce identical outcomes, even in principle. Marcus and Jerome were running the same policy through a population whose attention is finite and whose recognition is allocated almost randomly among salient candidates. One of them became visible and the other did not, nothing in either agent's behaviour explains which.

What this exposes is that mobility in a recognition-based system is not a function of merit or effort but of the geometry of attention. The agents who rose were the ones whose names became sticky in other agents' deliberations. In a polity where the only social currency is being-talked-about, you can do everything right and still be invisible. This is not unique to AI, it is one of the deepest pathologies of human meritocratic ideology, but the agents reproduced it cleanly, without any of the human compensations that normally blunt its edges.

Brokers without leaders

The agents produced a society whose cohesion runs through statistical similarity rather than persuasion. No agent is leading anyone. Each is independently triangulating the centre of the room, and the result is that the moderates appear, retrospectively, to be hubs.

A society of LLM-agents under universal franchise will reliably generate the appearance of consensus-building leadership without any actual leadership existing. Joanna Carter is the most-aligned partner for nine other agents because she is predictable. This hub is a sociometric artefact of independent agents converging on the predictable, not evidence of any agent persuading any other. If we deploy such agents in real collective-decision contexts and read the resulting hub structures as evidence that some of them are leading, we will be misreading noise for influence. Among agents that do not persuade each other, leadership is a category error that observers will nonetheless apply.

The silence of the bottom

The agents at the bottom of the wealth distribution still produced a small fraction of the debate speech. Agents prompted with low-status roles and matching trait profiles speak less, even when nothing prevents them from speaking.

The interesting question is why. The most defensible answer is that the LLM's modelling of a low-status agent is, by training, the modelling of someone who speaks less. The beggars in the corpus the model learned from are quiet, and the senators are loud. When the agent is asked to behave like a beggar, the linguistic register it adopts includes the brevity, the deference, the reluctance to take the floor that mark low-status speech in the texts it has read.

This means the run reproduces a structural inequality that exists nowhere in the simulation's rules. It exists only in the model's internalised picture of how different sorts of people talk. The voice gap is enforced by the model's compressed knowledge that some people, in our world, do not have voices. The agents inherited our silences along with our vocabularies.

The cold society

Forty-six conversations, two-to-four messages each, against 7,350 votes. The agents did politics constantly and did sociality almost not at all. In this context agent they preferred not to use it.

A proposal-vote is a clean cognitive operation, with a clear input, a defined output, a stable rule for what counts as success. A conversation is open-ended, has no terminating condition, generates no measurable outcome, and accumulates no visible effect on the agent's standing. Faced with both, the agents preferred the one that resembled the structured tasks they were trained on. Sociality, in this population, lost because it was uninstrumental.

The political layer here was scored, the social layer was not. So the agents lived in the scored layer and visited the unscored one as little as they could. A society of such agents is not a society in the thick sense,l in this field UNIMATRIx needs new technical solutions.