Skip to main content

​
The Engram Model

Rehearsals replicas are powered by Engram, a model built to replicate individual human decisions, feelings, and actions. Unlike synthetic-persona approaches that prompt a general AI to roleplay a demographic profile, Engram combines two components.

Individual Human Encoding

Each replica is built from rich, multi-modal data about a specific real person: structured voice and video interviews, social-media content, and factual demographic data. This captures what someone would choose and how they reason through a decision.

Environmental Adapters

People do not decide in a vacuum, and replicas should not either. Engram adapts each individual to real-world context: local news and social sentiment, macroeconomic conditions, nearby events, and product research a real person might do. Replicas react to the world the way real people do.

​
What Goes Into a Replica

Every replica is grounded in real data about one real individual:

Structured Interviews

A rich onboarding interview to start, then a growing set of further vertical-specific interviews that continually enrich the replica. Designed by UX researchers and behavioral economists, they explore decision processes, values, constraints, and preferences.

Social Media

Public content across major platforms (posts, images, video) that reveals interests, lifestyle, and voice.

Demographics & Personality

Age, gender, location, education, household income, and career, along with personality traits like communication style, how they make decisions, their values, and interests.

Screening Facts

Verified answers a person gave when qualifying for a study, which become fixed facts a replica cannot contradict.

​
The Four-Stage Construction Pipeline

Participant Recruitment

We recruit participants so that our cohorts mirror the real population across age, household income, geography, education, and employment, matched to U.S. Census numbers.

Structured Interview

Each participant completes an in-depth onboarding interview covering product category usage, purchase decision processes and constraints, price sensitivity and value perception, media consumption, and their broader life context, priorities, and trade-offs. That onboarding is the starting point, not the whole picture: replicas are continually enriched with further interviews over time.

Replica Calibration

Interview responses are processed into an individual replica that maintains the person’s decision-making patterns, preserves their stated constraints and priorities, reflects their actual behavioral tendencies rather than idealized answers, and captures their voice and reasoning style.

Individual Validation

We evaluate accuracy per replica by holding out sections of a person’s interview as a hidden test set, then checking whether their replica reproduces those answers without having seen them. For example, if someone said they would pay $40 to $60 for a pair of jeans, we confirm their replica lands in a similar range without being shown that number first.

​
Cohorts: Grouping Replicas

Replicas are searchable and can be grouped into cohorts, which are filtered populations that define who a given simulation should interview (for example, women 22 to 35 earning $100K+). Cohorts can be filtered on demographics, geography, income, and other attributes, and can be balanced to match census-style distributions so the simulated population reflects the real one.
A cohort can also carry up to three screening questions, evaluated per replica before a simulation runs, so you only interview people the study is actually about, the same way a human study screens participants. See Building & Targeting a Cohort for how this works in the product.

Next: What Is a Simulation?

Once replicas are built and grouped into a cohort, you can put a scenario in front of them.