Jefferson is a generative-agent simulation of the electorate, in the lineage of Park et al.'s generative agents and Argyle et al.'s silicon sampling. It is not a "world model" in the ML sense — it doesn't learn environment dynamics. It simulates a population, and it is built around one organizing thesis:
Cross-sectional fidelity does not transfer to interventional fidelity. A simulator can match every poll and still be useless for planning — many causal models produce the same observed totals.
So Jefferson is two products on two timelines. Now-casting — "is my picture of the electorate right now correct, by segment?" — validated against polls and election results, shippable today. The intervention engine — "what does this message actually do, and to whom?" — validated against randomized message tests and ad holdouts, the research frontier. We never conflate the two, and neither should you.
┌─────────────────────────────────────────────┐
│ PLANNER · external RL [R&D] │
│ which intervention moves the outcome? │
│ optimizes ON the population, never IN it │
└──────────────────┬──────────────────────────┘
▼ acts on
┌──────────────────────────────────────────────────────────────────┐
│ THE WORLD (always changing) news · ads · local events · polls │
│ ▼ ingested per precinct │
│ ENVIRONMENT STATE what this electorate was exposed to + │
│ believes now — exposure routing + salience filter │
│ ▼ │
│ AGENT POPULATION grounded per precinct · pattern completion │
│ + social diffusion — agents follow social rules, │
│ they do not maximize rewards │
│ ▼ ▼ │
│ VALIDATE vs fresh polls SERVE prompt → instant segmented │
│ → drift-gated recalibration race snapshot [TODAY] │
└──────────────────────────────────────────────────────────────────┘
Each precinct's population is synthesized with iterative proportional fitting over Census/ACS marginals, joined with voter-file aggregates (turnout history bands, party registration shares) and attitude priors from public survey waves (ANES, CES) plus licensed omnibus data. The result: a population whose joint distributions match the precinct, where no agent corresponds to an identifiable person.
Why this matters twice: it's the privacy posture and the statistical posture. We calibrate to distributions, not individuals — the same foundation that makes MRP the modern standard for small-area opinion estimation.
The most important design split in Jefferson: agents follow social rules; the planner optimizes. Voters in Jefferson behave by predictive pattern completion — "what does someone like me do in a situation like this?" — with memory, reflection, and response. There is no scalar reward inside an agent, because there is no scalar reward inside a voter. This follows the Theory of Appropriateness line of work rather than rational-choice agent design.
Opinion doesn't form in isolation booths. An idea spreads because people observe it and update their sense of what's normal — not by cost-benefit calculation. Scaled to a precinct, repeated observation produces cascades: tipping points, subcultures, influence networks. Recent multi-agent work shows this engine generates realistic macro phenomena — Veblen effects, influencer-driven subcultures — without hard-coded utility functions. Swap "luxury good" for "political position" and you have the closest thing to a persuasion mechanism currently available.
A poll measures opinion at t₀. Jefferson's roadmap models the trajectory: where a message lands, through whom it spreads via household and proximity ties, where it dies — consistent with complex-contagion research showing social behaviors spread through clustered ties, not broadcast exposure. We say "roadmap" deliberately: diffusion parameters are hard to identify from observational data, and we won't sell a cascade we haven't calibrated.
Simulators drift. Jefferson's answer is a standing loop: sparse ground truth from cheap polling (within TCPA and state-law constraints), backtests against precinct-level election returns, and drift-gated recalibration — retraining triggers on measured divergence, never on a clock, and only against trusted data. The trap we engineer against: auto-tuning to noisy, house-biased polls teaches a system to be wrong, confidently.
The keystone dependency is honest to name: every autonomous loop terminates at "get a fresh, trusted, representative sentiment reading cheaply." That is a human-data-collection problem, not a compute problem — and it's where much of our engineering effort goes.
Step one, already underway: a single-precinct historical backtest. Pick a past race with a known result, synthesize the precinct's population, poll it, turnout-weight, and measure the gap — overall and by subgroup, because a model can be right on average and wrong in every subgroup. Every later layer is validated against this ruler. The sequence only unlocks in order.
The campaign-facing layer has two parts. The insights agent — patterned on the internal data-agent architectures published by frontier labs — plans queries over simulation runs and campaign data, executes them, verifies its own numbers, and answers in plain language. Message testing, ad testing, and segmentation are conversations, not SQL.
The planner is where reinforcement learning belongs — and the only place: an external optimizer that tries interventions against the simulated population and learns which ones move outcomes. Its reward is the simulator's own output, which is why it ships behind the calibration discipline, with conservative offline-RL penalties in low-data regions and ensemble disagreement as uncertainty.
The trap we build against: optimize a planner against an uncalibrated simulator and you don't discover what moves real people — you discover what exploits your simulator's biases. Reward-hacking your own model. Calibration gates everything.
Park, O'Brien, Cai, Morris, Liang, Bernstein (UIST 2023, Best Paper). Introduced the memory-stream / reflection / planning architecture for believable agent behavior — the architectural backbone of Jefferson's L2.
arxiv.org/abs/2304.03442 →Park, Zou, Shaw, Hill, Cai, Morris, Willer, Liang, Bernstein (2024). Agents built from 2-hour interviews replicated participants' General Social Survey responses at 85% of the accuracy with which participants replicated themselves two weeks later — the existence proof that calibrated agents can stand in for people on survey instruments.
arxiv.org/abs/2411.10109 →Argyle, Busby, Fulda, Gubler, Rytting, Wingate (Political Analysis, 2023). Demonstrated "algorithmic fidelity": conditioned on demographic backstories, LLMs reproduce response distributions of human subgroups on political surveys — the statistical basis for L1's persona conditioning.
doi.org/10.1017/pan.2023.2 →Centola & Macy (American Journal of Sociology, 2007); Centola (Science, 2010). Behaviors requiring social reinforcement spread through clustered networks, not just weak ties — why Jefferson models household and proximity edges explicitly rather than treating exposure as broadcast.
doi.org/10.1086/521848 →Fine-tuning an open-weight model on ~2.9M responses across 210 social-science studies to predict experimental outcomes. Open weights + fine-tuning + treatment-effect targets — the closest published recipe to Jefferson's intervention engine, and the source of the honest failure numbers below.
arxiv.org (SocSci210) →Leibo et al. (Google DeepMind). Individuals generate behavior by predictive pattern completion — accounting for norms without scalar rewards or preference relations. The strongest argument against making voter-agents reward-maximizers, and the reason Jefferson's agents follow social rules while only the external planner optimizes.
arxiv.org →Paglieri, Cross, Cunningham, Leibo, Vezhnevets (DeepMind). Evolved persona-generating programs that span what's possible in a population (support coverage), not just what's probable. Jefferson uses coverage to guarantee the tails exist, then re-weights to real demographic and turnout proportions — conflating the two is the mistake to avoid.
arxiv.org →Cross et al. (DeepMind, Concordia lineage). Status symbols, Veblen effects, and influencer-driven subcultures emerge endogenously from social observation + imitation — no hard-coded utility. A working model of preference shift through social diffusion: the candidate persuasion mechanism behind Jefferson's L3.
arxiv.org →Park, Gelman, Bafumi (Political Analysis, 2004) and successors. State- and district-level opinion from national samples via multilevel regression and poststratification — the statistical standard Jefferson's synthetic-population calibration extends to precinct scale.
doi.org/10.1093/pan/mph024 →Bond, Fariss, Jones, Kramer, Marlow, Settle, Fowler (Nature, 2012). Social messages on Facebook measurably changed real-world turnout, with effects flowing through close ties — field evidence that the diffusion channel Jefferson models actually moves votes.
doi.org/10.1038/nature11421 →The field has validated replication of outputs, not a mechanism. The current scorecard: simulated average treatment effects correlate ~0.85 with real experiments — but effect sizes are overestimated 2–10×, effect direction is wrong 10–32% of the time, and subgroup variation gets flattened. Those failure modes are exactly Jefferson's research agenda: get the sign right on subgroups first, calibrate magnitude second, and report distributional fidelity (not just point estimates) throughout.
That's why every Jefferson claim links to the evaluation run that produced it, why we publish calibration error alongside predictions, and why simulations are decision support — they tell you where to spend your next real-world dollar, not what to believe instead of reality.