Simulated audiences and THYMOS
Every result in Thesia comes from simulated people. This page covers where they come from, how they decide, and how far to trust what they tell you.
Who is in an audience
A simulated person, called a persona in the API, is a generated profile: age, occupation, location, interests, goals, pain points and habits. A group of them is an audience. You can get one in four ways:
- Describe the segment in plain language, for example "owners of 5 to 20 person agencies in Bucharest".
- Give a market, so the people reflect that country.
- Give your website, so they are grounded in what you sell.
- On Studio and Pleroma, import your own customer records.
Manage them under Audiences in Context. Once an audience exists, you can use it again across runs and compare results on the same people.
How a decision gets made
THYMOS is our cognition model. It holds the state a person decides from, such as how they appraised what they saw, how much effort it has cost them and how confident they are, and carries that state forward. A language model on its own tends to produce a reasonable answer. THYMOS is there so the choice comes from state and not from whatever sounds most plausible.
- A language model reads what the person was shown and appraises it from their point of view.
- THYMOS updates their state and chooses what they do next.
- A language model puts that choice into words.
Because the state carries over, two people from the same segment can reach the same screen and act differently.
Which runs use it
| Run | How the reaction is produced |
|---|---|
| Read | A few language-model appraisals, then the THYMOS choice model over a weighted population. Takes seconds. |
| Site audit | THYMOS decides at each step while the person moves through the page in a real browser. |
| Audience conversation | THYMOS decides the person's stance, and a language model writes the reply. You can open their inner reasoning. |
| Study, creative comparison, post simulation | A language model produces each person's reaction from their profile and what they were shown. |
How far to trust it
Treat a result as a prediction to act on and then check, not as a measurement. A read is labelled as a simulated prior for that reason. A simulation is most useful for finding what confuses people, ranking options and surfacing objections early. It is weaker on exact percentages, and on anything your audience knows that the model does not.
You can test this on your own work. Seal a prediction before you launch, which locks it with a hash and a timestamp, then record the real outcome against it afterwards.
How many people to use
The method sets a sensible default: 30 for a message or concept test, 40 for pricing, 8 for a focus group or interviews, and 3 for a site audit. More people cost proportionally more. One person's answer is an anecdote, so run a study before you make a decision on it.
Next: Understanding usage · All docs