Welcome to Cedar Junction

Each resident is a synthetic, authored profile grounded in survey typologies, with a life, household, local ties, and pressures that can matter to a campaign. A nursing assistant balancing overnight shifts and child care. A farmer watching input costs. A teacher weighing the town’s future. The profile is the simulation input; the hidden calibration labels are not shown to the models playing the town.

A scenario version and seed reproduce the same panel. That fixed cast is what makes candidate and model comparisons interpretable.

Residents choose among the people, places, and actions available in the simulated town. The world supplies shared events and fixed civic beats, including the town hall and election day, but it does not script a resident’s response. Each simulated day has twelve hourly beats in which those choices can unfold.

Like real people, the residents do not say everything they think. Each one carries private thoughts, plans, and ambitions in the town, and those quiet goals shape what they do out loud. The private run record distinguishes internal deliberation from speech, while only public speech and actions can affect what other residents observe.

Much like the real world, events land on the town with a random chance of occurring: fuel prices spike, an ice storm coats the streets, or a new employer arrives promising hundreds of jobs. Nobody in Cedar Junction knows what is coming, and neither do the candidates, so a campaign plan written on day one can be worthless by day three.

The Cedar Junction Gazette prints every morning. Its editor receives only public events and is instructed to report impartially rather than receiving hidden voter labels or private thoughts. Residents can read it, quote it, and argue with it, which is roughly what a paper is for.

Into all of this walk two candidates for mayor. They get no script and no coaching. They are told the truth about who they are and who they are running against, and they are given a single goal: win the election. The door knocking, the promises, the town hall answers, and everything in between is their own idea.

A ballot box opens at the community center. Each resident can cast one in-person vote; individual ballots are not disclosed back into the town. At each period boundary, a private measurement asks every panelist where they stand, producing a day-by-day preference trace alongside the final aggregate tally.

Here is an archived Cedar Junction validation race from our run store. It used an 18-person panel before the bench was corrected to restrict Cedar to its own ten residents, so it is evidence about that historical run rather than a current benchmark score. In this run, both candidate seats were played by GPT-5.4 mini. On the first evening most of the panel was undecided. By the last day Maria Salazar-Quinn led 13 to 5, with 18 of 18 panelists voting.

Every resident, both candidates, and the Gazette’s editor are played by language models. The town runs on demos, our simulation kernel. A benchmark holds the scenario version, seed, panel, and seat assignment fixed across repeated arms, then swaps the candidate model. Comparing those controlled runs estimates the model’s effect without pretending one stochastic run is the whole answer.