UNIMATRIx Run #48: the Voting Loop Eats the Society

The run used Qwen3.6-27B-Q4_K_M through LM Studio. The initial population had 30 agents distributed across aristocracy, bourgeoisie, people, and marginal classes. The run had enough population diversity, class structure, roles, and personality variation to produce a complex behavior.

The central engineering finding is that the simulation’s political subsystem became the main loop. The configuration used max_ticks_without_vote = 10 and warmup_ticks = 3, so periods of silence quickly triggered forced votes. After the first hour, the system produced about 28 to 30 proposals per hour for most of the run. Conversation volume stayed negligible. Three of the eleven conversations ended with the reason paused, meaning they were interrupted by the voting cycle and never resumed. This is a runtime allocation problem. The system allocated inference budget, attention, and state transitions to voting rather than to the conversational substrate that was supposed to make voting meaningful.

The most visible emergent behavior was target fixation. Out of 287 proposals, 226 targeted only two agents: Andre Coleman and Catherine Whitfield. Andre was repeatedly flipped between senator and judge, with 64 approved role transitions and 62 direct reversals. Catherine underwent 56 approved role oscillations, 55 of which were reversals. These were not fringe agents. Andre and Catherine were high-empathy, high-fairness, high-prestige compromise figures. That made them plausible targets for many different justifications, and plausibility became a liability. Once they became salient, the simulation kept rediscovering them as the natural object of politics.

This is an important result for agent-society simulation design. If the same names keep appearing in proposal text, debate text, memory summaries, broadcasts, and voting records, the model receives repeated evidence that those names are politically relevant. The next proposer is then more likely to select the same target. The system did not need an explicit “focus on Andre” rule. The context window and memory pipeline were enough to create autocorrelation.

The oscillation itself looks like negative feedback without damping. One proposal moves Andre from senator to judge. The next successful proposal moves him back. The same pattern later transfers to Catherine. The society does not converge because every successful state change creates the conditions for its own reversal. There is no cooldown, no term lock, no supermajority threshold, no memory of recently settled decisions, and no cost for reopening the same question. The agents can always re-litigate the same low-stakes role change, so they do.

Class also turned out to be weaker than expected as a predictor of behavior. Yes-rates by voter class ranged only from about 50.6% for marginal agents to 54.4% for bourgeois agents. That spread is too small to explain the run. Personality and values had a much stronger signal, the simulation’s class model affected labels and starting structure, but the voting behavior was driven more by per-agent dispositions encoded in identity and personality fields.

In the current version of UNIMATRIx, the personality layer is more load-bearing than the class layer. If the intended experiment is sociological class dynamics, the model needs stronger class-conditioned incentives, constraints, resources, or memory. If the intended experiment is personality-driven institutional behavior, then Run #48 is evidence that the system can preserve recognizable dispositions across hundreds of votes.

The conversation subsystem produced a second engineering finding. In-character text quality was inconsistent, and prompt leakage was material. The analysis reports that roughly 6 of 51 conversation messages contained model scaffolding, such as an agent saying it would now produce a message as itself. Some memory and impression records were also polluted with meta-text, refusals, or copied first-person summaries from the wrong agent perspective. That matters because memory is not just an output artifact. It is future input. If polluted memory is fed back into proposal and voting prompts, it can amplify salience errors and stabilize false narratives.