Definition
A study is the operational unit of synthetic research. It specifies what to ask, who to ask, and how to evaluate the resulting evidence.
Good studies prioritize comparability and reproducibility. Inputs are versioned, execution is tracked, and outputs are reviewed by cohort and aggregate views.
Operational model
- Design layer: define objective, population, and skill before execution.
- Execution layer: run prompts across population members with consistent context.
- Observation layer: collect response quality, sentiment, and divergence signals.
- Decision layer: convert findings into product or policy actions with traceability.
Execution workflow
1. Design
Document hypothesis, target population, success criteria, and constraints before first run.
2. Execute
Launch the study with stable inputs and monitor run health while outputs are collected.
3. Analyze
Break down results by cohort, inspect outliers, and map findings back to the original hypothesis.
4. Iterate
Adjust skill framing or population composition as a new study version, then rerun.
Quality checklist
- Study objective and decision owner are explicitly documented.
- Population and skill versions are fixed for the duration of a run.
- Run output includes both quantitative summaries and qualitative examples.
- Outliers and contradictory responses receive explicit review.
- Exported findings preserve enough context for later audit.
Failure modes to avoid
- Changing prompt wording during active runs and losing comparability.
- Running studies against poorly validated populations.
- Ignoring low-confidence or contradictory outputs because averages look stable.
- Skipping export/audit artifacts, making findings hard to defend later.
Related platform APIs
These APIs are the main building blocks used when this concept is put into production workflows.
/personas/promptRun skill execution for persona responses.
/personas/prompt/streamStream long-form study responses.
/personas/prompt/jobs/:jobIdObserve async execution status.
/personas/prompt/sessionsInspect run history and outputs.
/integrations/workflows/timeline/reconcileQueue post-run consistency checks.