Poll the simulated.
Persuade the real.
Test your message on simulated voters — indistinguishable from real ones — then go find the real ones.
Four things. Any order the race needs.
Test it before a real voter hears it.
Find the real ones, down to the street.
The poll people actually answer.
Send what worked.
Traditional polling is great. Until it sucks.
Simulated voters you can't tell from real ones.
One question at a time.
Real communications research — not just consultants.
None of the 2004 playbook.
The method is a trade secret.
Traditional polling is great. Until it sucks.
The classic phone poll reaches the ~3% of people still willing to answer an unknown number at dinnertime, weights that sliver until it resembles an electorate, and reports a margin of error as if sampling were the only thing that could go wrong. It isn't. It's the least of it.
Sub-5% pickup. The people who answer aren't a random slice of anyone.
"Likely voter" screens are educated guesses dressed up as arithmetic.
The margin of error is the smallest error in the poll. It's just the only one they can calculate.
We fixed the parts of silicon sampling everyone else is still tripping over.
Simulated respondents are only as good as the errors you've engineered out of them — the caricatures, the agree-with-everything reflex, the flat little distributions that nod too politely. We spent a long time on exactly those. We solved them. We could walk you through how — but that would rather defeat the purpose of a trade secret, wouldn't it.
Layers most people never model
Certified returns. Crossover proofs down to the household. Donor behavior. Consumer and communications signals — layers most polling never touches.
Real communications research — not just consultants
We don't stop at political scientists and campaign consultants. We build on persuasion science, psycholinguistics, and message framing — the research on how people actually change their minds.
The silicon-sampling errors, solved
Mode collapse, sycophancy, caricature, over-agreement — the known failure modes of AI-simulated respondents are the ones we engineered out. How, exactly? Trade secret. That's rather the whole point.
We skipped the 2004 playbook
No random-digit dialing. No dinner-hour call banks. No eight-part "generic ballot." Much of the field still treats that playbook as canon — defended with the energy Star Wars fans reserve for what counts as canon.
One question. One idea. Wild concept.
Most polls staple two or three issues into one question and call the mush "data." Nobody can agree with a paragraph — so a "yes" tells you nothing. We ask one thing at a time. It's slower. It's also the only version that means anything.
"Do you support the Governor's plan to cut taxes and increase funding for public schools?"
"Do you support increasing state funding for public schools?"
No poll is perfect. Ours is just wrong less — and never on live TV.
We built the locked room, then lost our own key.
Every client gets a sealed workspace. Row-Level Security scopes each query to your race and your geography — enforced by the database, so even we can't wander into your voter file by accident.
app_reader can read your data and write nothing to the engine. Ever. That's not a promise, it's a permission.
An assistant locked in the vault with your data.
Ask the map anything in plain English and layer on your own canvass data. It works only inside your sealed workspace — read-only, sandboxed to your race, and logged. It sees everything you're scoped to and changes none of it.
// look, don't touch.