Psychohistory for Pocket Change: Building Voxari Studio

In Isaac Asimov's Foundation, Hari Seldon invents psychohistory, a branch of mathematics that cannot tell you what one person will do but can tell you, with uncomfortable precision, what a galaxy of them will do. It only works at enormous scale. Seldon spends his life on the equations, records his conclusions in a vault, and dies long before anyone can check whether he got them right.
It is one of the most seductive ideas in science fiction. It also has a flaw that every data scientist will recognize: nobody could audit Hari Seldon. No Per Review.
A few weeks ago, I wrote here about Voxari, a living world of 5,000 synthetic citizens where Claude does the thinking and Jev, TypeSafe's judging model, does the judging. Voxari is free, public and a little bit silly on purpose. This article is about what grew out of it: Voxari Studio, my attempt at a small psychohistory that shows its work, prints its own error bars, and costs less than the coffee you are drinking.
In summary: it is a platform where you can create a simulation of a population and run studies on it. What population? Any, The USA, Europe, your company, a hospital, or even Gotham City and Hogwarts.
Ask a synthetic population before you ask a real one
That is the line on the front page of studio.voxari.org, and it is the whole product in one sentence.
You describe a group of people. Studio builds them, asks your questions, and shows who leans which way, by how much, and why. The group can be the 2,000 employees of a logistics company, the residents of Georgia county by county, the nurses of a hospital, the members of a union, or the students of Hogwarts.
To create a population, is quite simple. You can create one from scratch, define your sample size, add traits, trait values and proportion. Simple for a simple population, like a 50-people company, but not for something more complex like a whole state, county by county, with dozens of traits. To help with that, I created templates, over 200 hundreds of them and counting.

Every invented person gets traits drawn from a specification: role, tenure, age, pay band, shift, county, education, household income, modelled party lean, whatever the template carries. Then Jev model reads each person and judges how they would answer. The rule from Voxari survived intact: Jev judges; it never writes. At most 16 people per request, a handful of short questions, and every answer comes back as probabilities with a confidence value attached. No essays, no invented quotes from people who do not exist.
The unit of usage is the response: one person answering one question. When I measured it, a response cost me about $0.000006. All 5,000 Voxari citizens answering 3 questions came to about nine cents. A 10,000-person study with nine questions runs around 55 cents. Hari Seldon needed the budget of an empire. This needs loose change.

Where the people come from
One engine, many templates. A census state, a hospital and a fictional world are different specifications run by the same code.
The workplace and community templates (company, hospital staff, hospital patients, customers, university students, members of an association) ship with illustrative distributions that I wrote. You are expected to replace them with your own numbers. You can upload shares per department or per site, but never rows of real people: no employee lists, no patient records. The upload parser rejects anything that looks like a list of individuals.
The real places are where most of the work went. Georgia was first, and it set the pattern:
Pull county-level tables from the American Community Survey (age by sex, race and ethnicity, education, income, household type, employment).
Pull the Census microdata records for the state with the same variables.
Adjust the microdata weights, using iterative proportional fitting, so that every one of the 159 counties matches its own totals.
Sample each invented person from the weighted records of their county, so every person is a realistic combination of traits and not a random mix of independent columns.
Add the modelled traits and label them as modelled: party lean from county election returns (MIT Election Lab) and religion from the US Religion Census.
Check that every county's generated people land within tolerance of the census, and write the error report next to the template.
From there it spread to every US state by county, Brazil by state and by município (IBGE's 2022 Census), the United Kingdom, and the European Union by country and by region.

Small areas need care. At 10,000 people, one invented Georgian stands for about 1,100 real ones, and Wyoming gets around 17 seats out of 10,000 on the national template. Any group with fewer than 30 people is greyed out on the results page, with the reason. I would rather show a grey box than a confident number built on five people.
And yes, there are fiction templates: Middle-earth, Tatooine, Westeros, Hogwarts, Panem and Arrakis, among others. They are unofficial, they name their authors and studios, and they exist because teaching statistics with Panem's twelve districts is much more fun than teaching it with "Region A" and "Region B".
My first question was a bad one
The first live study asked a company population: "We will change our health insurance coverage. How do you feel about it?"
The result was 100% undecided, with Jev 90% confident about it.
My first reaction was that something was broken. It wasn't. Nothing in the question said what would change, so no trait could make anyone win or lose. A driver with two kids and a manager with no dependents had exactly the same information: none. Jev was right, it was my question that lacked the needed information.
That failure shaped three features.
Context. A population can carry a background (the situation everyone shares) and each study can carry a briefing (what people know before answering). Jev judges from both.
The question coach. Before a run, Claude Opus reads each draft question and flags the classic sins: vague, two questions in one, leading, loaded words, missing context, jargon. It proposes up to three complete rewrites, a briefing, and up to three traits the population is missing that the answers would depend on. This is the one place where Studio uses a model that writes, and its words reach Jev only if the author picks them. Each check costs three to five cents. When it doesn't know a fact, it writes a placeholder like [amount], and Studio refuses to run a question that still has brackets in it.
Robustness checks. A question can carry up to three neutral rephrasings. Studio splits the people between the wordings, so the check costs no extra responses, and then reports how much the answers moved. Under five points the result holds, under ten it moves somewhat, above that it is fragile. If "which wording they saw" explains more of the answers than any trait of the people, you have learned something important about your question, not about your population.

How wrong is it?
It is not a big deal to be wrong, as long as you know why and how to be less wrong. That was a lot of the work. I built a standard gold record: real, published survey questions from Pew, Gallup, AP-NORC, KFF, the General Social Survey, the Cooperative Election Study and Quaest, none about candidates or parties, each with its source, dates, wording, sample size and group results. Before running anything, I set aside every third question as a held-out set, chosen by hash so I couldn't cherry-pick. All tuning happened on the rest.
The measure is the total difference: the share of people who would have to change their answer for the synthetic result to match the real one. For scale, a second real survey of the same size would land about 1 point away.
The first reading, counting each person's most likely answer, was 26.8 points off on the held-out questions. Then came the cheapest improvement I have ever shipped: instead of counting picks, average each person's probabilities. That brought it to 21.1, at zero cost. Counting picks throws away the close calls. A person who is 60-40 gets counted as fully one way, and a thousand of those make every result look much tidier than people really are.
Some things did not work, and I think they deserve as much space as the things that did. That is how we learn: more from our mistakes than from anything else.
Jev avoids the extremes of a scale. Asked whether they are "extremely proud" to be American, 33% of real respondents said yes; Jev gave 3%. In Brazil, 71% of people said violence had "increased a lot"; Jev said 14%. I tried a mathematical correction that pushed weight toward the ends of every scale. On the tuning questions it looked wonderful. On the held-out questions it fixed some results and wrecked others: where real people sit in the middle (most think it's "probably false" that mRNA vaccines alter your DNA), the correction destroyed an answer Jev had nearly right. The earlier, smaller test had looked like a clear win, which is exactly why I didn't trust it. It looked a lot like overfitting.
Jev doesn't know recent local policy. In Brazil, a question about ending the escala 6x1 (six days of work, one day off) landed 34.7 points away. Jev reads Portuguese fine; it just didn't know what the 6x1 schedule is. A two-sentence briefing explaining it brought the gap down to 9.4. The coach now warns about this on every question that names a recent law or a local term.
Then I asked Claude Opus 5.5 to read the same invented people and answer the same questions. On the held-out set it landed 7.8 points off, against Jev's 21.1.
My first suspicion was memory: maybe Claude had simply read those polls during training. So, I added a fresh batch of seven questions fielded from July 2026 onward, after Claude's training data, and neither model could have seen the answers. Both got worse, which is honest: Jev went to 33.9 points off, Claude to 15.9. But Claude was still closer on 6 of the 7. The one miss was a Pew question on data centers and the environment, where Claude saw more harm than people do.

In the last article, I leaned on Daniel Kahneman's System 1 and System 2. Studio ended up with the same split. Jev is System 1: fast, cheap, and it reads every single person. Claude is System 2: slower, several times dearer, so it reads only a sample of 50 at a time. Then Studio moves Jev's answers toward Claude's, group by group, along whichever trait drives the question most. With those 50 people, the adjusted result lands within a point or two of what Claude alone would give with 200. On the more capable plans this happens automatically on every run, labelled Adjusted with Claude (Beta), for about 15 cents a check (my cost). Jev's own answers stay right below it, so you can see the disagreement.
The number I put on the result front page is the most important one for me because it shapes what this tool is about: on 20 published questions, four answers in five landed within ±22 points of the real survey. That is a wide margin. Studio is much better at contrasts between groups than at exact percentages. It can tell you that caregivers will hate a policy more than everyone else; I wouldn't bet my job on whether 35% or 41% of the whole company opposes it.
It is great to show you direction, but the real numbers depend on how good you are in describing context and writing hypothesis.
Even the trade-off studies were checked. I ran two published conjoint experiments from the political science literature through Studio and compared. On an immigration study, 12 of 13 effects that were clearly different from zero came out with the right sign, and the ranking of what matters most was mostly right. The sizes came out about half of the real ones. Direction good, magnitude timid, which is Jev's strength and limitation.
What a result looks like
Three finished studies are public, and they are also the three guided studies every new account gets for free. They work as a tutorial to the simulation system.

A logistics company moving everyone to a high-deductible health plan. 20% support, 35% oppose, 45% undecided. The split that matters is caregiving: people caring for children or a relative sit at a net lean of -38, those without at +6. If I were the HR director, that is the most important finding.
Ending the twice-yearly clock change, across the US by state. 47% would keep standard time all year, 31% daylight saving time, 5% would keep changing. Claude, asked about 50 of the same people, came out 22 points away. The report shows that disagreement, so you can have your interpretation.
A Ministry inspector arrives at Hogwarts. Slytherins lean +18, Gryffindors -23. Then you hold the other traits constant and the house barely matters: trust in the Ministry explains 88% of the variation. Sorting Hat, you have been overrated.
Time, forks and futures
A population, as generated, has no memory. Scenarios give it one.
You write an event (a title, what happened, who it reaches) and every person reached is asked how it leaves them and what it does to their trust in those responsible. Then you advance a month, which is free and instant: word spreads inside each group (department, county, party lean), people who only heard about it take on part of their group's reaction, and those who lived it drift toward their neighbours.
Then you fork. One branch where the company announces the new health plan in January, another where it announces it together with a raise. Run the same study on both, and Studio compares them question by question, telling you which differences are larger than the noise.
This is the part that feels closest to Seldon's vault, except that what comes out is a rehearsal of two possible futures, and you get to pick which one to live in.
Where I think this goes
I built Studio because I kept seeing the same expensive pattern: an organization spends weeks and real money on a survey, and half the questions turn out to be badly worded, or aimed at the wrong group, or answerable with common sense. A synthetic population is a cheap first pass that tells you which questions deserve the expensive one.
Some places where I think it fits:
HR and internal communication. Rehearse a policy before announcing it. Find the group that will feel it most and the wording that backfires.
Public health and city departments. There are templates for health, schools and public safety in Brazil, the US and Europe, and a county-level map is exactly the kind of thing these teams need for messaging.
Market research triage. Run twelve questions on Studio, find the three with real tension between groups, and field those to real people.
Teaching. Studio tests hypotheses, applies the Holm correction when you test several, and shows error bars that include both sampling and model error. Doing that on a population of Hogwarts students or the districts of Panem makes a statistics class much easier to stay awake in.
Writers and game designers. Ask Tatooine how it feels about the Empire, then send in an event and advance six months.
Agents. Team plans get a REST API and an MCP server, so Claude (or any other agent) can build a population, run a study and read the results without a human clicking anything. Given how much of my work now runs through MCP, this is the feature I am most excited about.
There are also partners: companies that resell Studio in reais, with a nota fiscal, and run studies for clients who want answers rather than software.
And there are things Studio will not do, by design. It will never accept records of real people. The synthetic notice on every report, export and share card cannot be removed, not even on white-label plans. In Brazil, where electoral law requires polls to be registered before publication, published reports on geographic templates do not name candidates or parties in election years. I would rather lose a customer than have a Studio chart passed around as a real poll.
Is it free? For a lot of features, yes. However, I could not make it entirely free for one reason. This project and the others I run, like two games (The Hollow Dream and Vesperfall), one Mind Engine and Simulation, a recomendation system, and more, cost real money that paid customers will help to fund my innovation iniciatives.

Real Tests
A few weeks ago, Brazil had their presidential election. In the first round, far-right wing candidate Flavio Bolsonaro was ahead of incubent president Lula. Since neither candidate reached the mininum 50% of valid votes (disconsidering abstention, blanks and nulls), they had about two weeks until the run off election between only the two of them. Every poll predicted Bolsonaro to win, I decided to to put Voxari Studio to the test and created a poll on a 10,000 sample of the Brazilian simulation. The result of what I got was really surprising.
In my study, Lula would finish first. Considering only votes that are valid, he would end up with close to 52%. That would mean he would have to do what had never happened in Brazilian elections. A candidate in 2nd in the first round of a presidential election had never came ahead on the run off. It seemed odd, but I built Voxari and I decided to take the risk. I publish the study, locked in the date and recorded a video on my social media.
The result? By now you already know. In a historical come back, Lula was elected. The world was in shock, so was I. You can double check this study in here. And Yes, studies can be made public and shared.
Candidate | Voxari Studio | Real Result |
Lula | 51.19% | 51.79% |
Flávio Bolsonaro | 48.81% | 48.21% |
Results, considering only valid votes
Candidate | Voxari Studio | Real Result |
Lula | 43% | 40.48% |
Flávio Bolsonaro | 41% | 37.69% |
Blank or Null | 6% | 3.24% |
Did not vote | 10% | 19.22% |
Results, considering all votes
Seldon, revisited
Hari Seldon's great trick was that nobody could check his math until it was too late to matter. He appeared in the vault, delivered his verdict, and vanished.
Studio does the opposite. Every share comes with its error bar. Every report says the answers come from a simulation. The benchmark tells you how far it lands from real surveys, and Claude sits beside Jev pointing out where they disagree. It is a much less romantic psychohistory, and that is the only version I would be comfortable selling.
If you want to try it, go to studio.voxari.org. Three guided studies are on me, no card needed, and there is a Feedback link on every page. Tell me what broke, what confused you, and where the invented people got you wrong. That is the whole point of a psychohistory you can audit.




Comments