Audience Construction

AnthroSim builds each synthetic profile from a population-weighted U.S. demographic cell and attributes decoded by the installed Fusion model. Before a run, the audience checker labels which criteria the current data can apply and which remain qualitative guidance.

Example input: 10 profiles · Audience group: “commuters affected by a proposed rail extension”

Each saved profile records:

• The applied criteria: supported age, education, income, occupation, and survey-derived filters
• The audience guidance: the label and notes supplied by the operator
• The generated person: name, background, attitudes, and communication style used in the room
• The run context: model artifact and segment definition stored with the transcript

Illustrative synthetic profile. It does not describe a real person.

What the research supports

In AnthroSim’s first study, calibrated profiles reduced aggregate error on tested ANES political outcomes to one third of naive repeated prompting. Several cheaper methods performed about as well on simple population readout. The later individual study found that invented traits did not improve predictions about a named person. Measured facts about that person carried the useful signal.

Profiles therefore create controlled variation inside a synthetic room. Individual predictions require measured facts and separate validation. Deep demographic descriptions also cannot recover a variable that the underlying survey never collected.

Read the population study →

What the run exports

Every saved run exports the generated profiles alongside the transcript, operator-visible private traces, final ballots, and source-use records in JSON.

Next: Conversation simulation →