Frequently asked questions

Candor is a synthetic research platform for deciding what to build, ship, and price with evidence, not instinct. It builds synthetic users grounded in real evidence, combining your uploaded research (when you have it), web data, and validated behavioral models to give each one realistic personality traits, cognitive biases, and decision-making patterns. Use it for problem discovery, problem validation, concept testing (including name, value proposition and messaging tests) and price testing, on new products and live ones. You get decision-grade signal the same day, where a panel study would take weeks.

Most AI persona tools generate fictional characters from a prompt. Candor starts with evidence (your documents, web research, validated behavioral data) and builds each synthetic participant from the ground up. Each of the fields describing who a participant is, how they behave, what they believe and how they talk carries a provenance tag saying what is behind it (personality and bias scores, and the demographics summary at the top of the profile, carry none): grounded in something you can check, derived from that evidence, or assumed and flagged as such. A critic agent validates every interview response for consistency before you see it.

Candor is built for the teams that own research demand. UX research leaders use it to scale their research capacity when their team can run one study a week and the PMs need five. Customer and consumer insights leaders use it to pressure-test concepts and pricing without burning panel budget. Innovation teams use it when a stage gate is approaching and traditional research would take too long. Consultants and agencies drop it into client engagements as a fast-validation layer. Startups use it to test assumptions before committing engineering resources. You don't need to be a trained researcher to run a study.

Candor uses a multi-stage pipeline, and the order matters more than people expect. It reads your audience description and any uploaded documents, then proposes candidate segments along behavioral and psychographic dimensions rather than demographics alone. Only then does it search the web, gathering evidence about those specific segments and extracting signals as it goes. That is why a sharp audience description changes the whole result: it is the input to the first real decision the pipeline makes. From the settled segments it builds groups, samples individual synthetic users with OCEAN traits and cognitive biases drawn from real population distributions, and writes narrative memory for each one.

Evidence-grounded means a synthetic participant is built from evidence gathered for your audience before anyone is generated, and the fields describing who they are, how they behave, what they believe and how they talk are each labelled with what is behind it. Grounded means there is something you can check: evidence found for your audience, a value you set when you set the study up, or U.S. Census population data, and hovering the label says which. Derived means it follows from that evidence without a source that states it. Assumed means Candor filled it in and says so, so you know what to trust and what to probe. Personality and bias scores, and the demographics summary at the top of the profile, carry no label.

Provenance tagging is Candor's system for recording what is behind each field on a synthetic participant. Each of the fields describing who the participant is, how they behave, what they believe and how they talk carries one of three tags. Grounded means there is something you can check: evidence found for your audience, something you told us when you set the study up, or public population data. Derived means it follows from that evidence without a source that states it. Assumed means nothing pointed at it and the model says so. So for any participant, you can see which fields rest on something you can check and which don't. Personality and bias scores, and the demographics summary at the top of the profile, carry no tag.

Candor offers two interview modes. In live mode, you type questions and the synthetic participant responds in real time. It feels like a natural conversation. In auto-interview mode, you configure an interview guide with sections and learning goals, and Candor conducts the interview automatically, tracking coverage and adjusting questions as it goes. Both modes use the same inhabitation system, where the model takes on the participant's full profile, and a critic agent validates every response for consistency.

Candor synthesises every transcript into a report, automatically, as soon as the last interview finishes. The pipeline runs a different sequence depending on the interview type you chose, because a hypothesis verdict and a willingness-to-pay range are not built the same way. What you read is consistent: four sections, opening with what matters most, which holds the learnings, then the findings, then the next steps, then an evidence appendix holding the supporting detail and the full transcripts. In problem validation, concept testing and price testing each learning is paired with the move it implies; problem discovery is for understanding, so its learnings carry no action; its top move still sits in what matters most, and the rest are in the next steps. Today's reports paraphrase the evidence rather than quoting it, and the transcripts are there when you want the exact words.

Each synthetic participant is built on Big Five (OCEAN) personality traits sampled from peer-reviewed population distributions, calibrated by region and occupation. Cognitive biases are assigned with research-backed intensity values. Every participant has narrative memory covering their identity, behaviors, beliefs, and communication style. They remember everything across interview sessions, maintain consistent views, and push back when they disagree. Each one has a set skepticism level that controls how often they push back, to counter the underlying models' tendency to agree.

Every response a synthetic participant gives passes through a critic validation step before you see it. The critic checks for hard contradictions (the participant said something that directly conflicts with an established belief or prior statement) and soft tensions (a notable shift in enthusiasm without a clear reason). Hard contradictions trigger regeneration. Soft tensions are flagged in metadata but still delivered, because real people have variable moods too.

Candor uses the Big Five (OCEAN) model for personality: Openness, Conscientiousness, Extraversion, Agreeableness, and Neuroticism. Trait values are sampled from cross-cultural population distributions published in peer-reviewed research. Cognitive biases are modeled as first-class traits with intensity values from 0 to 100, drawing from established behavioral science. B2B and B2C synthetic participants use different attribute models reflecting distinct decision-making frameworks.

Yes, for the fields describing who they are, how they behave, what they believe and how they talk. Each one carries a provenance tag, so in the participant detail view you can see which are grounded in something you can check, which were derived from that evidence, and which are assumed, and hovering a tag says what it rests on. The profile doesn't name the document or signal behind a field, and personality and bias scores, and the demographics summary at the top of the profile, carry no tag. Synthesis reports don't carry these tags. Instead, discovery findings show how many participants raised them, a new validation report shows how many participants agreed with each hypothesis you tested, out of those asked about it, and names the ones who disagreed (one built before this was added counts everyone who spoke to the topic), validation and concept-test reports name the participant behind each paraphrased piece of evidence, and every report comes with the full interview transcripts.

Four interview types, and concept testing splits into four kinds of its own. Problem discovery explores unknown problems in a market. Problem validation tests whether your hypothesized problems are real. Concept testing evaluates a product concept: what resonates, what needs rethinking, and who it's for. Price testing examines value perception gaps, anchoring effects, and willingness-to-pay ranges. Inside concept testing you pick a concept test type. A standard test shows whole concepts cold and compares them. A name test, a value proposition test, and a messaging test each expose one anchor concept first and then present the candidate names, benefits, or headlines inside that context, so nobody judges a name in a vacuum. See the use cases page for the full set.

Yes. You can upload PDF, DOCX, CSV, or TXT files: customer interviews, survey data, market reports, or any research you have. Candor parses, chunks, and indexes these documents, then uses the facts and quotes they state as the primary evidence source for audience generation. Those are prioritized over web research, and each signal extracted from them names the file it came from on the audience page. Where a document's author interprets their own evidence, Candor records that sentence as the author's reading rather than as a fact: it shows as Derived, and it cannot on its own define a group or a dimension. A participant's profile doesn't name documents: the fields describing who they are, how they behave, what they believe and how they talk are labelled grounded, derived or assumed instead (personality and bias scores, and the demographics summary at the top of the profile, carry none).

About an hour or two, most of it in the background. Your audience and its synthetic participants generate together in one run of 25 to 35 minutes, and it is queued, so you can close the tab. Interview guides generate in a couple of minutes. Auto-interviews run 25 to 35 minutes in parallel across your participants, and live interviews happen in real time. Synthesis adds 3 to 5 minutes once the last interview finishes. Your own attention adds up to well under an hour of that, spent describing the audience and reading the report. You go from a research question to a structured report the same morning, instead of the six weeks a panel-driven concept test would take.

Yes, and it treats them as fundamentally different. A procurement lead evaluating enterprise software uses a different decision framework than a consumer making an impulse purchase. Candor models B2B and B2C synthetic participants with distinct attribute schemas, personality weightings, bias profiles, and buying triggers. You choose your audience type when creating a study, and the entire pipeline adapts accordingly.

Your data is stored in Supabase (hosted on AWS in the US) and is never shared with other customers. Uploaded documents are used only for your studies. Data processing happens through secure APIs with industry-standard encryption in transit and at rest. Candor's full list of third-party data processors is published on the subprocessors page, updated whenever vendors change.

No. Your research data, uploaded documents, and interview transcripts are not used to train any AI models. Candor uses commercial AI APIs for inference only. Your data is processed and returned, not retained for model training. This applies to all third-party AI providers in the stack.

After every response a synthetic participant gives, a separate AI model reviews it against that participant's established beliefs, prior statements, seed decisions, and stance change history. It checks for hard contradictions (directly conflicting with something the participant said or believes) and soft tensions (a shift in tone or enthusiasm without a clear reason). Hard contradictions are rejected and regenerated. Soft tensions are flagged but delivered, because real people have variable moods.

Candor samples Big Five personality traits from peer-reviewed cross-cultural research, including Schmitt et al. T-score distributions across 56 nations and Anni et al. occupational personality profiles. Regional and occupational distributions are used when available. When occupational data doesn't exist for a specific role (about 57% of occupations), Candor falls back to regional norms. No traits are assigned randomly.

When you upload research documents, Candor parses, chunks, and indexes them. During audience generation, each signal extracted from your documents carries a reference to the source file. A fact or quote the document states is tagged 'grounded'; a sentence that is the document author's own reading of their evidence is kept as a hypothesis and shows as 'derived'. This provenance carries through to each synthetic participant, whose fields describing who they are, how they behave, what they believe and how they talk are labelled grounded, derived or assumed (personality and bias scores, and the demographics summary at the top of the profile, carry none), and to their interview behavior. A synthesis report doesn't link findings to documents; you check a finding against the interview transcripts that come with it.

Synthetic research is not a substitute for talking to real people. It's a complement. Candor's evidence grounding, personality calibration, and consistency enforcement make its output more structured and traceable than prompting a generic AI. But synthetic participants can't replicate genuine emotional responses or surface truly unexpected insights the way a real conversation can. Use Candor to sharpen your hypotheses and identify where to invest real research budget.

Sign up at runcandor.com. Every new workspace starts with a trial project, so you can build an audience, generate synthetic participants, and see a real report before choosing a plan. You create a study by describing your target audience, optionally uploading research documents, and choosing your study type. From there, Candor handles audience generation, synthetic participant creation, and interview facilitation.

Yes. You can sign up today with a work email. Start on the Free plan with no card, or choose Standard. The pricing page compares the two. Candor is built by Highline Beta, based in Toronto, Canada.