How a study works

A study asks one question about one product: how would these people react to it? Jahan answers by building a population of simulated people, letting them meet the product and each other over simulated time, asking them whether they would buy it, and recording everything that happened so every number can be checked.

This page walks through a study from start to finish. Each step has its own page with the detail.

flowchart LR
  B["<b>Brief</b><br/>product · price · claims<br/>audiences · assumptions"] --> P["<b>Population</b><br/>personas drawn from real<br/>respondents + a social network"]
  P --> S["<b>Scenarios</b><br/>variant × price × channels"]
  S --> W["<b>Worlds</b><br/>one scenario × one seed"]
  W -->|"every tick"| C["<b>Channels</b><br/>feed · forum · word of mouth"]
  C --> V["<b>Survey waves</b><br/>purchase intent, in words"]
  V --> W
  W --> T[("<b>Trace</b><br/>every turn,<br/>whole ticks")]
  T --> R["<b>Report</b><br/>digests · findings<br/>objections · trust level"]

1. You write a brief

The brief describes the product: its name, category, price and competitors, the claims it makes ("20g protein with zero sugar"), where each claim came from, and the audiences you want to study ("people who train three or more times a week, 60% of the sample"). Anything you take as true without evidence is written down as an assumption, and every report repeats it.

2. Jahan draws a population

Every simulated person, a persona, is one real row of a persona dataset: a survey respondent, a product reviewer, a public figure. Personas are sampled, never invented. The draw is checked by statistical gates before any model is called; a draw that does not look like the audience you asked for is refused.

The personas are then tied into a social network shaped by who resembles whom, and the network's communities are discovered in it.

3. Scenarios are run as worlds

A scenario is one version of the product at one price, reaching people through a chosen set of channels for a stated number of ticks (hours, days or weeks). Each scenario is run several times with different seeds; each run is a world. How much the worlds disagree is the study's estimate of its own noise.

4. Personas meet the product and each other

On every tick, some personas are active. Each active persona is shown what reached it on each channel: an X-like feed, a Reddit-like forum, or a friend telling it about the product. It reacts in its own words: it likes, replies, posts, tells a friend, or ignores. Its beliefs and memories carry forward to the next tick.

5. Survey waves measure intent

At scheduled ticks every persona is asked how likely are you to purchase the product? and answers in its own words. The answer is turned into a five-point distribution by comparing its meaning to fixed reference statements, never by asking a model for a number.

6. Everything is recorded

Each persona's reaction is one turn, and every turn is written to an append-only trace, one whole tick at a time. A study can be stopped, resumed and replayed exactly. The prompt behind any turn can be rebuilt from the record and is shown only if it matches the fingerprint taken when the turn ran.

7. The report is derived from the record

Results are computed from the trace and nothing else: adoption per wave, how far beliefs moved, the objections people raised in their own words, and patterns such as herding. Every finding carries the trace records behind it and the real-world test that would prove it wrong. Every report states its trust level, which today is always uncalibrated. See Limitations for what that means.

A study at a glance

Step	Input	What happens	Output
Brief	your YAML file + a category ontology	checked against the engine's contracts	a validated brief and its assumption ledger
Population	corpus + study size + seed	filter, sample, gate, fill gaps, build network	personas, network, communities, gate report
Worlds	scenarios × seeds	ticks of activation, exposure, reaction	the trace
Analysis	the trace	digests, objection clusters, anomalies, findings	the report (Markdown + JSON)