The science
Published science. Not prompt engineering.
Every synthetic identity the Market Thinker works with is born from validated personality models and national census data. The language model enters only after the identity already exists.
The unit
A synthetic identity represents a region of a population, not a person.
To test your alternatives, the Market Thinker uses synthetic populations. Each one is born from the Population Genome: a country's recipe, built from official census data, with no personal data. Within it, Synthetic Identity Models describe regions of the population, each a center of personality, motivation and culture plus its natural variation, at four levels of resolution: national, regional, cultural and tribe, a tribe being a small, tightly knit group around a passion. From these models, synthetic identities are created on demand to react to each alternative. A synthetic identity is not a Thinker and not a twin: a twin is a 1:1 replica of a real person, built from their data and with their consent.
Population Genome
The recipe of a country: census cells with their weights, personality norms and culture.
Identity Model
A region of a population — national, regional, cultural or tribe. A center plus its variation.
Synthetic identity
An instance of an Identity Model, created on demand for a simulation. A copy of no one.
Twin, something else
A 1:1 replica of a real person, built from their data with their consent. Synthetic populations are not made of twins.
The computational-expressive boundary
The math decides. The language model gives it a voice.
The language model is responsible only for expression. It gives voice to an identity that already has opinions, preferences and a way of deciding.
What a synthetic identity believes and prefers is determined by the scientific framework before the language model is involved. Demographics, personality, culture, market behavior: all computed. Zero AI.
This separation is what produces culturally specific, coherent answers instead of the statistical center of what a language model was trained on.
Coverage, not N
Why a market does not need one identity per person.
Studies with people need large samples because what drives each person cannot be observed. In a synthetic population it is explicit. So the job changes from estimating a hidden distribution to covering a known one.
Copies add nothing
Give a language model the same profile and the same question, and it gives essentially the same answer. A thousand copies of one identity carry the information of one.
Variety is the sample
What adds information is a different identity. Effort goes where a population is most diverse; homogeneous groups need few synthetic identities, mixed ones need more.
Size enters as weight
How many people an identity stands for is a weight taken from the census, not a headcount to be simulated one by one.
To understand a market, you need coverage of its identities, not copies of its people.
The blind spot
Language models don't understand culture. We encode it.
A persona written in a prompt flattens a whole country into a stereotype. “A 35-year-old Brazilian woman”, told to a language model, comes back as an American default with a Brazilian label.
Same country, different worlds
Brazil has Afro-Brazilian communities in Bahia, Japanese-descended families in São Paulo, German-heritage towns in the South and evangelical movements reshaping urban peripheries. A single “Brazilian persona” captures none of this.
Culture shapes decisions
The same product draws opposite reactions from cultural groups inside one country. Our cultural layer captures those differences as measured differences in personality and values, not as stereotypes.
Statistics, not stereotypes
Cultural groups come from demographic data and published normative studies. Traits vary together with ethnicity, region and religion, the way they do in the real population.
“A synthetic Nikkei-Brazilian and a synthetic Nordestino don’t just have different names. They decide differently.”
Architecture
An eleven-layer cognitive genome.
Eleven independent layers, from the census foundation to memory. Each adds a dimension that a prompt alone cannot supply.
Census Demographics
Official national data: gender, age, socioeconomic level, education, region, urbanization, ethnicity. The mathematical foundation.
FoundationIdentity Models
Personality, motivation and decision style, organized into Synthetic Identity Models at four levels: national, regional, cultural and tribe. Not a label — a statistical description of a region of the population, with its natural variation.
CalibratedCultural Genome
Diversity inside a country, encoded as measured differences between cultural groups.
DifferentiatorMarket Behavior
Industry attributes across finance, health, food, media, retail, transport and values — data, not a language model’s memory.
Industry dataHousehold & Life Context
Family structure, household size, marital status, children, income. Decisions happen at home, under real constraints.
ContextDecision Mode
Some decisions are made on intuition, others by deliberate analysis. Each synthetic identity approaches each decision in the mode that fits who it is and what is at stake. Kahneman, operationalized.
Decision scienceBehavioral Style
How a synthetic identity behaves in a group or in a professional role. The same person acts differently in different rooms.
Context-dependentMigration & Lifestyle
Language at home, acculturation (Berry framework), mobility inside and across countries. A professional from São Paulo living in Miami gradually shifts lifestyle, consumption and decisions.
AcculturationProprietary Enrichment
Your customer data and your own segments, added as a versioned layer you can roll back. It never contaminates the shared genome.
Client-ownedScenarios
Populations calibrated to a described future. “If 30% of your customers adopted GLP-1 and stopped drinking.” “If your competitor launched this.” A/B testing at population scale.
What-ifMemory & Evolution
A synthetic identity can keep a memory of the simulations it joins and evolve with them, anchored to the region it represents. It never invents a life inconsistent with its genome.
Evolution11:11 is an alignment — the moment when census data, personality science, cultural encoding, decision theory, behavioral modeling, migration patterns, proprietary intelligence, scenario engineering and persistent memory converge into something that could not have existed before this moment in technology.
Our own evidence
What we measured, against real people.
Our Behavioral Validation Experiments compare synthetic answers with human answers to the same questions, across language models and conditions. Correlations are synthetic versus human.
Demographics alone. Age, gender and income do not reproduce how people answer.
Personality alone, in stage 1 of the study: decisions and affinities tied to personality traits. In the later stages, with operational behaviors, fidelity drops sharply (see below).
of the variance in results comes from how the question is designed.
comes from the choice of language model. The remaining 20.8% is the interaction between the two.
Where it is weaker, stated plainly. Language models reproduce who people are better than what they do day to day: correlation of 0.318 for identity attributes against 0.121 for operational behaviors, under the same condition. That is why we read results the way we do.
How to read a result
A filter, not an oracle.
Order
Which option wins, and which comes last.
Direction
Whether a change moves people toward or away from you.
Difference
Which groups react differently, and why they say so.
Not exact percentages. A simulation narrows the options worth taking to real customers; it does not replace them.
Scientific foundation
Standing on published research.
Three independent research programs support the principles our architecture puts into practice.
Stanford University
Park et al., 2024Generative agents grounded in rich individual profiles reproduced people’s answers 85% as accurately as those people reproduced their own answers two weeks later.
The principle we build on: the richer the grounding of each identity, the closer the simulation.
Nature (Binz et al.)
Centaur — a foundation model of human cognition, 2025Trained on 10.6 million human choices from 160 psychology experiments, it outperforms domain-specific cognitive models. Its acknowledged limitation: a strong bias toward Western, educated populations.
A census-based, culturally calibrated population is the piece that work says is missing.
PyMC Labs & Colgate-Palmolive
Maier et al., 2025 — 57 consumer studies, 9,300 peopleLanguage models reproduced human purchase intent at 90% of the maximum achievable correlation, but only with rich demographic conditioning. Without it, the figure fell to 50%.
Grounding is not optional. It is the method.
Positioning
What we are not.
Not agent-based modeling
Traditional agent-based models give agents behavioral rules and calibrate at the aggregate level, by matching macro outputs. Synthetic identities are defined by empirically calibrated profiles, which allows opinions, reasoning and stated preferences alongside quantitative distributions.
Not a language model used as a simulator
The academic literature (Argyle et al., Santurkar et al., Törnberg et al.) documents both the potential and the limits of asking a language model to play a population: training-data bias, underrepresented minorities, inconsistency within the same agent. Those are the problems our architecture addresses, by separating what is believed from how it is said.
Validation
Accuracy is not claimed. It is earned.
The five tests of our validation protocol, and two observations from the field.
Control
Synthetic results against real-world behavioral benchmarks.
Consistency
The same specification gives statistically equivalent results.
Discrimination
Different segments give measurably different answers.
Calibration
Stated confidence matches observed hit rates.
Sensitivity
Changing a trait shifts behavior in the expected direction.
Peru — Datum Internacional
Synthetic Peruvians gave answers that researchers with deep in-country expertise recognized as culturally authentic, including patterns of household decision-making, family structure and emotional register that language-model-only approaches had systematically misrepresented.
In partnership with Datum Internacional. Full documentation in progress.Brazil — a crisis in real time
While a major global sportswear brand faced a public crisis in Brazil, our synthetic population, with no exposure to the real-world reaction, surfaced the same negative signals: emotional rejection, brand dissonance and cultural misalignment across several consumer segments.
The synthetic signal and the real crisis ran in parallel, neither informing the other. Formal documentation in progress.Scope & commitments. We are transparent about what this framework is and is not. Ethnographic work, the discovery of needs people cannot articulate, and genuinely new cultural phenomena remain domains where human expertise is irreplaceable. Our framework is a complement, not a substitute. No real individual’s data is used without appropriate consent. Census data is public. Published science is cited. Outputs are clearly identified as synthetic, never presented as human-generated data.
Coverage by country
The base of every simulation, country by country.
Every simulation the Market Thinker runs starts from a country's Population Genome. This is where those genomes stand today.
Census + Identity Models at every level. United States, Brazil, Mexico, Colombia, Peru.
Census cells ready, Identity Models in progress. Canada, Poland, Ecuador, Portugal, Paraguay and more.
Census and personality studies gathered, genome build on the way.
live, across the countries with a full genome.
“The same precision used in climate models and epidemiological forecasting, applied to understanding how people think, choose and act.”
Thinkers.pro Science Team