A simulation asks a group of digital twins to make one decision, then reports where they landed. Five fields define it, and the quality of the answer depends almost entirely on how carefully you fill them in.
The five fields
- 01PopulationWho is being asked. Pick one you have already built, or describe a new one and let the agent assemble it.
- 02DateThe moment the scenario is set in. This is not the date you run it. Asking about August 2026 and August 2019 are different questions, and twins reason from what was true then.
- 03ScenarioThe situation each twin is placed in. Concrete beats abstract: 'deciding a tip on DoorDash after ordering' produces sharper reasoning than 'thinking about tipping'.
- 04QuestionWhat each twin is actually asked, in the second person, as a person would hear it.
- 05Action spaceEvery option a twin may resolve to. At least two. Twins must end on one of them, so leaving out a real option forces an artificial answer.
Writing an action space that holds up
The most common way to get a misleading result is an incomplete action space. If the honest answer for part of your population is “I would not order at all”, and that option is missing, those twins are pushed into a choice they would not make and your percentages shift.
If you are not sure the space is complete, run it once with a catch-all option. If the catch-all takes a meaningful share, you are missing a real branch.
Reading the result
A finished run gives you a distribution across the action space, a confidence level, and the individual twins behind it. Start with the twins. The percentages tell you what happened; the transcripts tell you why, and the why is what survives contact with a decision.
Confidence is measured, not asserted. It reflects how closely this population has matched held-out humans on questions it had not seen. Treat a high-confidence run as reliable for direction and ranking, and the exact percentages as approximate.
Changing one field at a time
The comparison view is the point of the tool. Run the same population and question against two dates, or two scenarios, and the difference is attributable. Change several fields at once and it is not.