Start from what is known.
Every attribute keeps its origin.
Public statistics, electoral records, survey microdata and geographic data are ingested separately, mapped to one schema and kept with their source, date and level of aggregation attached. Each component is refreshed on its own cycle as new evidence is published.
No source is used raw or alone. Each is weighted against the others, so agreement strengthens a signal and disagreement flags it.
Census & official statistics
Population structure, households, education, employment and material conditions, published at the smallest available geography.
Age structure differs sharply between the two localities. Each synthetic individual is drawn to match the structure of its own geography, never the national average.
Electoral records
Published results at polling-district level, treated as the observed behaviour of a place — never as an attribute of any individual.
The spread inside each district matters as much as its mean. Two places with the same headline result can hold very different mixes of neighbourhoods.
Survey microdata
Whole respondent profiles that preserve how characteristics co-occur — the joint structure a table of averages loses.
Income and education travel together, but not perfectly. Sampling whole profiles keeps the off-diagonal people — the graduate on a low income — in the population.
Geographic & administrative data
Boundaries, settlement patterns, infrastructure and service access, joined to every locality in the graph.
Distance from a service is a property of place that shapes how a message about that service is received. It is attached to the locality — and so to everyone placed there.
Public opinion & media
Published opinion measurement and information-environment indicators, kept with their date so context is never mistaken for a current reading.
Media diets shift over time. Each indicator carries its publication date, so an agent’s information environment reflects the period being simulated.
Illustrative figures. The charts show the shape of each source, not a current reading.
Define what things mean.
Before anything is counted.
A formal definition of the entities and the relations between them — what a household, a locality, an institution or an affiliation means, and how each relates to the rest. A dataset stores attributes side by side; the ontology says how they hang together.
Seven dimensions
ARDA holds the meaning of each attribute and the relationships between them: what a neighbourhood, an income bracket, a union, a church or a news diet imply about each other. That is what lets a population be constructed rather than assembled.
Connect the evidence.
Six layers, one representation.
The evidence instantiated as a connected graph, so context is retrieved through relationships rather than looked up in a row. Six layers carry it from ingestion to validation — and grounding and reasoning are kept apart by design at every one of them.
Sources cleaned, mapped to one schema, with provenance kept on every observation.
An ontology and a knowledge graph. Not attributes stored side by side, but the relationships between geography, income, education, politics, media and culture.
A synthetic population sampled from ARDA to match your target audience, preserving the combinations of characteristics that actually occur together.
Every agent reads the same stimulus from its own state and context. Outputs are structured: interpretation, sentiment, behavioural intention.
Responses aggregated at population and segment level — distributions, where the audience splits, distance from the intended outcome.
Predictions benchmarked against observed outcomes wherever comparable evidence exists; the gaps feed back into the population model.
Build whole people.
And keep them whole.
Attributes are not invented one at a time and stapled together. Wherever the evidence allows, a synthetic individual is built from a whole observed respondent profile, so the combinations of characteristics that actually occur together stay together.
Observed respondent records provide the joint structure — the way income, education, geography, media use and belief co-occur in real lives.
The population is weighted and calibrated against official population structures, so its margins match the geography and audience you selected.
Each individual is placed in its context — locality, institutions, information environment — through the knowledge graph rather than a written description.
No personal data is used or produced. Runa models populations, not identifiable individuals.
Hand each agent a person it did not choose.
Then let it read.
ARDA decides who the agents are and what context they have; the language model decides only how each of them processes what you show it. Every agent reads the same message from its own position and returns a structured response: what it understood, how it felt, what it would do next.
One agent, one reading
The model never chooses who it is speaking as. It receives a person and a context assembled from the graph, reads the stimulus from there, and returns a response the pipeline can aggregate — across 400 agents, by segment, against your intent. What comes back →