Reading a participant profile

Profile, interview, and history

A participant has three pages. The detail page is the profile, and it carries two buttons at the top right that open the other two: Interview [name] 1:1 opens the live interview page, and See previous interviews opens the session-history page, where every past session expands to its transcript. Under the name sits a one-line descriptor of who this person is, on participants whose profile carries the longer nutshell summary; on older participants that line is the nutshell card itself, so it appears once rather than twice. The breadcrumb trail at the top of each page takes you back. Everything below describes the profile page.

Demographics

A two-column grid at the top with the participant’s identity anchors. Axes vary by audience type. B2C participants typically show age, location, occupation, household, and education. B2B participants show role, seniority, function, company size, industry, and decision authority. If a heritage profile was generated (cultural, regional, or ethnic context that shapes language and decision style), it appears here too.

Cognitive biases

A full-width section, open by default, listing every bias active in this participant sorted by intensity (highest first). Each row is a bar from 0 to 100 with a plain-English reading to its right: how hard the bias pulls (mild, moderate, strong) and what it does, for example Strong pull: picks the safe option nobody gets blamed for choosing. Above the bars, a short note on what biases are and a line naming the one or two biases the participant is most influenced by stay visible even when the section is collapsed. Biases aren’t labels; they’re calibrated intensities. A participant with loss aversion 78 reacts to potential losses far more strongly than one at 32, and that difference shows up consistently across interview turns.

The evidence decided which biases matter for the participant’s group, the model set each one’s range (when three or more groups carry that bias, a second pass sets it if the model keeps missing the spread between them), and the participant’s exact point was drawn from inside that range, not chosen by the model. Treat the list as a prediction: this is what you should expect to see in this participant’s reasoning.

Personality (OCEAN)

A full-width section, collapsed by default. Expanded, it opens with a short note on what OCEAN is and isn’t, then five rows scored 0 to 100 against real population data. Each row carries a reading of how the participant comes across in an interview at that score, for example Leans high (66): engages with hypotheticals and unfamiliar ideas instead of dismissing them. The five traits are:

Openness: curiosity, willingness to try new things, abstract thinking.
Conscientiousness: organisation, reliability, planning vs. spontaneity.
Extraversion: social energy, assertiveness, talkativeness.
Agreeableness: cooperation, empathy, willingness to defer.
Neuroticism: emotional reactivity, anxiety, stress response.

These aren’t random. Each value is sampled from peer-reviewed population distributions calibrated by region and, where applicable, occupation. Two participants in the same group will sit at different points within the group’s OCEAN ranges, which is why they sound different in interviews even when they share a worldview.

B2C or B2B profile

A collapsible section with audience-type-specific attributes. B2C participants have fields like life stage, financial pressure, digital fluency, brand sensitivity. B2B participants have fields like buying-center role, internal stakeholders, procurement constraints, evaluation criteria. Each field carries a provenance badge so you can tell where the value came from at a glance.

The four memory types

This is the heart of the profile. Participants have four memory types, each capturing a different aspect of how the person thinks. They persist across interview sessions inside the same project, so the participant genuinely remembers what you’ve discussed.

Identity memory: who they are

Background, role context, environment, and demographic anchors. The narrative version of who this person is and the situation they’re embedded in. Short paragraphs and bullets, not CVs.

Behavioral memory: what they do

Goals (what they’re trying to achieve), habits (recurring behaviors), constraints (what they can’t change), tradeoffs (the choices they’ve actively made), and usage patterns (how they engage with relevant tools or services). This is where the day-to-day texture lives.

Belief memory: what they think

Opinions (positions on specific topics, with the stance noted), objections (what would make them push back), preferences (specific aspect-by-aspect leanings), and decision heuristics (the rules they apply when choosing). Belief memory is what makes a participant disagree with you in interview rather than nodding along.

Language memory and interview behavior: how they talk

Vocabulary register, typical phrases, preferred tone, communication style, and verbosity target (terse, conversational, expansive, or narrative). This is what makes a participant sound like a specific person rather than generic AI prose. It also tells the model how long to make answers and which words to reach for. The last field, interview behavior, describes how this participant behaves in conversation: their tendency to elaborate, push back, ask clarifying questions, or get specific. Useful to skim before starting a live interview so you know what to expect.

Provenance badges

Every field in the memory sections and the B2C/B2B profile carries a provenance badge. The three values:

Grounded (green). There is something you can check behind it, and hovering the badge says which. The field cites signals in your evidence that, checked against the field’s own wording, match it (the hover says how many of its citations do; on a participant whose citations were never checked, a cited field takes its word from the model’s own tag). Or you fixed the value yourself when you set the study up, whether in a box like Age range or in your own words about who these people are. A box has to narrow something to count: “25-40” does, “any” reads the same as leaving it blank. Or it came from public population data: income, education and household on US consumer audiences, from the Census.
Derived (cyan). Built from your evidence, the audience you described, or population data, but no single source states this field’s value. Either fitted to the kind of participant your research described, where you did not set it yourself (where you left a setup box blank, Candor picked a value that fits the audience you described), or reasoned a step past what the research says outright: the evidence points here and may even be cited, but nothing in it states this value. Where Candor checks a field’s citations, each is tested against the field’s own wording, so related evidence that does not state the value reads Derived, whatever word the model chose.
Assumed (amber). The model is filling in the gaps based on everything it knows, but cannot trace this to any specific data points. Worth testing.

The point of provenance tagging is so you never confuse a confident claim with a guess. If a critical decision rule is tagged assumed, you know to test it explicitly in interview rather than treating it as established.

Evidence & provenance

Beside the demographics, a section titled Evidence & provenance summarises how much of the participant traces back to your research. Two main parts:

Sourcing: a colored chip, here and on each participant card, reading Well sourced, Partly sourced or Thinly sourced. It is counted from every badge on the page: add up the section headers below and you get the same numbers. Well sourced (green) means at most 15% of the fields read Assumed and at least half read Grounded. Thinly sourced (red) means more than 40% read Assumed, or fewer than 35% read Grounded. Partly sourced (yellow) is everything in between. A Derived field counts as sourced: it was built from your evidence or your setup, even though no single source states it. Hover the chip for the three counts and, when the participant is not Well sourced, for the fields that were Assumed. Because demographic fields count too, a study whose setup named an age or region has more fields reading Grounded than one that left those blank: your inputs are a source the badge can check.
What shaped this participant: three bars showing the share of fields reading Grounded, Derived and Assumed. Plus an evidence-depth line: average signals per trait, and the count of distinct signals used.

A participant whose citations were never checked (older studies whose sources were since deleted, or copies of another study) shows the same chip in grey, with no verdict, and the hover says so. Every participant is still interviewable whatever the chip reads. The chip tells you how much of a participant traces back to your research; it is not a to-do, because participants can’t be edited or regenerated one at a time.

How to actually read a profile

Three minutes is enough for a productive read:

Glance at the sourcing chip and the three bars first. That tells you how much to trust what follows.
Read identity memory and the top three biases. That gives you the person + the cognitive levers they pull.
Skim belief memory’s opinions and objections. These are what you’ll likely poke at in interview.
If you’re running an auto-interview, also skim language memory’s verbosity target and interview behavior. That’s what shapes how the participant will sound across the run.