Nothing here is a survey and nothing is a forecast from historical averages. Each twin reasons through the scenario on its own and commits to an action, and the distribution is what those individual commitments add up to. A run moves through five stages.
- 01SamplingDraw a weighted sample of twins from the population.
- 02GroundingLoad each twin with the signals that define it.
- 03ResolutionPlace each twin in the scenario and have it choose.
- 04AggregationCombine the choices into a distribution and segments.
- 05ScoringAttach a confidence level from past validation.
Sampling
A run does not use the whole population. It draws a sample, weighted to preserve the population’s composition, because a larger sample narrows the confidence interval with diminishing returns and rising cost. Two thousand is usually enough to separate options more than a few points apart.
The arithmetic is the same as for any sample. With two thousand twins, the margin of error on an option near half the vote is roughly two points either way. Quadrupling the sample only halves that margin, which is why larger runs are worth it for close calls and small segments, and rarely for the headline.
Grounding
Each sampled twin carries the signals that defined it: what it has bought, where it lives, what it follows, what it said in an interview. Those signals are what make one twin tip and its neighbour not, and they are the reason a twin’s stated reasoning can be checked against something.
Grounding is also what separates a twin from a model asked to “pretend to be a 34-year-old in Ohio”. A base model given only demographics tends to answer with the average opinion of people like that, which flattens real differences. A grounded twin answers from one person’s specific history, so two twins with the same age and income can, and often do, disagree.
Resolution
A twin is placed in the scenario at the specified date and asked the question. It may deliberate over several steps, but it must finish on one member of the action space. A run is complete when every sampled twin has resolved.
Forcing a single choice is deliberate. People rarely act on a “somewhat likely”; they buy or they do not. Asking twins to commit produces a distribution of actions rather than a spread of stated attitudes, and actions are what a launch or a price change actually meets.
Aggregation
The individual choices are counted, weighted back to the population’s composition, and broken down by segment. The reasoning each twin gave is kept alongside its choice, so any bar in the result can be opened to see the people inside it.
When a question has several scenarios, the same sampled twins answer all of them. That makes the comparison paired: a shift between two prices comes from the price, not from a different draw of people. Paired comparisons need far fewer twins to detect a real difference than two separate runs would.
Why the date matters
Twins reason from the world as it stood on the date you set. A question about delivery tipping in 2019 and the same question in 2026 reach different answers because the fee structure, the norms and the alternatives are different. Getting the date wrong is the quietest way to get a confidently wrong result.
What we do not claim
A simulation is a well-grounded estimate of how a group would behave, not a measurement of how it did. Use it to rank options, size a gap, and find the reasoning you had not considered. Do not use it as a substitute for the one real experiment you can actually afford to run.
- Rare eventsAn outcome chosen by one person in a thousand is below what a normal run can resolve. Do not read a small share as a precise rate.
- Market sizeA simulation says how a group splits between options. It does not say how many people will see your product in the first place.
- Situations nobody has facedThe further a scenario sits from anything the population was grounded in, the lower its confidence, and the result will say so.