Design a population

A population defines who faces the decision. A persona is a simulated respondent constructed for that population. Demographic selection, modeled traits, and a respondent's generated answers are different parts of the experiment.

Choose who is sampled

Start with one population mode. For a US adult starting point, use an explicit age range. With no state specified, the standard US draw uses twelve representative states; this is not a promise of a nationally weighted probability sample.

{
  "why_prompt": "How do price and driving range affect electric vehicle choice?",
  "experiment_type": "conjoint",
  "target_population": {"age": [18, 99]},
  "is_private": true
}

For a narrower US population, narrow the demographic constraints:

{
  "why_prompt": "How do price and driving range affect electric vehicle choice among adults aged 25 to 45?",
  "experiment_type": "conjoint",
  "target_population": {"age": [25, 45]},
  "is_private": true
}

Submit either complete example through Create an experiment. Review the question, design and cost before launch. These examples illustrate selection; they do not establish that the sampled adults are vehicle buyers.

Selection option	Meaning and constraint
target_population	US demographic selection. Age and income are inclusive ranges; other supported constraints use allowed-value lists. Supply an explicit constraint for the population you intend to study.
non_us_target_population	A separate non-US population mode. Coverage and constraint support depend on the population builder; use Holodeck to review the available configuration.
use_population_group and population_group	Select one named segment from the schema's allowed labels. A named segment is a selection rule, not a guarantee of identical individuals across runs.

Do not combine competing selection modes. For supported fields and allowed values, use the experiment request schema. For guided design and review before a paid launch, use Holodeck or the MCP workflow.

Add modeled traits

population_traits adds modeled characteristics to already selected respondents. It does not filter a population by observed behavior or prove that a respondent has a real-world characteristic.

Fetch the catalogue to read the trait descriptions:

curl "$SUBCONSCIOUS_API/api/v1/traits" \
  -H "Authorization: Bearer $SUBCONSCIOUS_TOKEN"

The experiment dictionary uses trait names and modeled values, rather than catalogue IDs alone. For example, add this field to either complete request above:

{"population_traits": {"Travel frequency": ["Weekly"]}}

This gives selected respondents modeled travel context. It does not identify verified weekly travelers. Always pair modeled traits with an explicit supported population selection.

Keep persona sources distinct

An uploaded population and external_personas are separate inputs. external_personas contains LinkedIn profile URLs when use_external_personas is enabled; it is not an arbitrary customer-data upload format. Holodeck's population upload uses a separate stored respondent file. Use the application's supported upload flow, and contact support@subconscious.ai about a first-party population before incorporating sensitive customer information.

Make comparisons reproducible

Record the selection rule, modeled traits, respondent count and seeds with the experiment. Holding population_seed and the configuration fixed with resample_population: false preserves the respondent draw under the same population data and implementation. Changing the source data or implementation can change that draw. Set resample_population: true when independent draws are part of the analysis plan.

A repeatable draw does not establish that the answers match human behavior. Keep the population definition alongside the design and validation evidence. See Research design and Human baselines.

Review before launch

• State who is included and excluded, and why that boundary matters to the decision.
• Separate observed demographic selection from modeled behavioral assumptions.
• Check the configured population and respondent count in Holodeck or the MCP draft.
• Keep selection and design fixed when the comparison is meant to isolate one change.
• Treat sample feasibility and external validity as separate questions.