Jefferson simulates your electorate as a living population of decision-making agents, precinct by precinct. Ask it anything a poll would tell you — by segment, on demand, validated against real results — for a fraction of the cost of a single wave. Then, on the roadmap: watch reactions spread through real social fabric before you spend.
A commissioned poll costs $20,000–$60,000, takes weeks to field, and tells you what a sample of people said at one moment — not why they said it, and not what they'll say after your message has moved through their households, group chats, and church parking lots. Most smaller campaigns can afford one or two photographs per cycle. Then they fly blind.
Each simulated voter is a generative agent with memory, reflection, and planning — the architecture behind research showing AI agents can replicate real individuals' survey responses with 85% of the accuracy of those same people repeating their own answers two weeks later. When your message reaches an agent, it doesn't look up a probability. It remembers, weighs, and decides — and the trace of that decision is yours to read.
Census, voter-file aggregates, and survey priors are fused into a statistically faithful population — every agent representative, no real person's data inside.
agents: 2,847 turnout_hist: matched party_reg: matched issue_priors: calibrated
Agents store what they've seen, form stances, and respond to stimuli through retrieval over their own history — attention, persuasion, conversation intent, turnout intent.
recall(msg, k=12)
reflect() → stance Δ
decide() → {attend,
persuade, share, vote}
Agents talk. Households, neighbors, weak ties — reactions propagate along a generated social graph, so you see the cascade, not just the first impression.
graph: household+geo+homophily cascade depth: 3.2 peak reach: day 9
Live toy render: one seeded message diffusing through a simulated precinct graph. In production runs, every lit node is an agent whose decision trace you can open.
Jefferson isn't a substitute for ever touching reality — it's a way to spend your contact budget where it matters. Compliant peer-to-peer text micro-polls provide sparse ground truth; precinct-level election returns provide backtests. Every calibration wave makes the simulation sharper, and every claim we show you links to the evaluation run that produced it.
The loop is the product: simulate, verify, recalibrate — a simulation that earns its predictions.
State house races. School board slates. Municipal measures. Primary challengers. The 99% of American campaigns that will never commission a tracking poll can now test messages, target persuadable segments, and forecast reaction — with the same class of tooling national campaigns pay seven figures to approximate.