Blind replication
Extra rows on a study you already ran.
Sample boosting with holdout validation. Send an instrument you already fielded and the audience cells you used. We return extra rows from a census-grounded panel of simulated Americans, same answer options your humans saw. You keep a human holdout and score us on a metric you freeze first.
Both happen on a 30-minute call. No account, no card. · Enterprise? Book a scoping call →
Or sign up with a work email and run it yourself →How it works
Three steps.
- 01
Send the study you already ran
Send the question stems, answer options, and the audience cells you fielded. Keep the holdout on your side; we never see those rows. We do not need respondent names or any personal data.
- 02
We match the cells and return rows
We match your cells on the panel and return simulated rows: same answer options your humans saw, same crosstab shape as far as your file supports. A person on our team maps your codebook by hand. This is Concierge or Enterprise work, not a self-serve upload.
- 03
You score the holdout
You score the holdout. We never grade our own work. Freeze the metric before we run (topline error, cell-level agreement, whatever you would use to judge a human sample vendor), then compare our rows to your human holdout.
Example results
What a run looks like.
You compute the miss
Predictions lock before we see your human results. The holdout stays on your side. That is the point.
Score us on a study you already ran →How we score the panel
650,000 census-grounded Americans with beliefs and persistent memory. Lewis, our proprietary model, runs them in a live simulation. Ask up to 10,000 at a time.
As shipped: about 10 pp on 1,521 real polls, 50 states. Calibrated benchmark: 7.47 pp on 404 non-electoral questions. Strict held-out: 9.97 pp.
Open-ended quotes and focus-group dialogue are qualitative and are not covered by these numbers. Election-margin questions are published but not claimed.
How we measure →What this is not
- Not a self-serve upload. There is no CSV-in, rows-out button here. A holdout rule and a frozen metric come first, and a person maps your codebook.
- Not Digital Twins. That product builds respondents from a written audience description and answers open-ended questions. Sample boosting matches the exact cells you fielded and answers the exact closed-ended items you ran.
- Not a benchmark guarantee. Our public 50-state validation compared a census draw against published polls with their own weighting. Your study is a different test: your cells, your questionnaire, your metric. That is the point of the holdout.
Hands-on on purpose. Codebooks, weights, and item screening need a human. Some items do not simulate well, and we will tell you which ones we will decline before we run. Concierge covers a single instrument. Enterprise covers a tracker.