Mode 1 is you typing questions and a synthetic participant answering in real time. It’s the right mode when you want to follow your instincts, push on something specific, or genuinely have a conversation. Here’s how it works.
Open any participant’s detail page from the participants list and click Interview [name] 1:1 at the top right. That opens the participant’s own live-interview page, where sessions start. Completed sessions live on another page, See previous interviews, opened from the button next to it. Each participant has its own session history; switching participants means switching contexts.
You type a question and hit Send message. Behind the scenes Candor does three things in sequence: it generates the participant’s response, runs a critic agent over that response for consistency, and then delivers it to you. You don’t see the intermediate generate/validate steps; you see the final turn with metadata attached.
Each delivered turn shows you:
The critic is a separate agent. Its job is to make sure the participant doesn’t drift mid-conversation. There are two severities. Hard contradiction: the response directly conflicts with an established belief or prior turn. These are caught and either regenerated or flagged so you can judge. Soft tension: the response is consistent but the tone or emphasis shifted in a way worth noting. These are tracked in the background but not surfaced inline.
The critic is what stops a Candor interview from devolving into agreeable AI prose. A participant who said earlier price isn’t the issue can’t turn around three turns later and say price is everything without either explaining why their stance changed or being flagged.
This is the part most people miss. The participant remembers the entire conversation, and they remember every prior session you ran with them inside the same project. End a session, come back tomorrow, start a new one, and they will refer back to what you discussed last time, including stance changes (which get logged explicitly, with a reason).
You can see prior session summaries from the participant detail page. They’re generated automatically when a session ends, distilling what was discussed and how the participant’s beliefs moved.
Memory alone isn’t enough to keep a participant consistent across many sessions, so Candor also keeps a ledger of the facts a participant has established about themselves. Their role, their circumstances, things they told you happened. Once a fact is in that ledger, it’s fixed. A participant who told you in session one that they manage a team of six cannot quietly become a solo contributor in session four.
Opinions are treated differently. Beliefs and preferences are allowed to move, but only with an explicit, logged reason. That’s the distinction to hold on to: contradictions in the facts are bugs, and shifts in the opinions are findings.
Participants have a snapshot of recent industry and region developments from the time the audience was generated. If you’re running a session weeks later and want a participant to know about something that’s changed in the industry, Refresh industry context rebuilds that snapshot. Use it sparingly. It pulls fresh evidence and recalibrates.
The button lives on the project page, not here. The interview panel shows the snapshot read-only, as a muted Industry Context line with a date; click the information icon beside it and a panel opens with the industry it focused on, the key themes it picked up, and links out to the sources it read. That date is the snapshot this particular session was frozen against, which is not always the newest one the project has.
A session freezes its snapshot the moment it is created, so a refresh never reaches a conversation already under way. If someone refreshes while you are mid-session, a yellow note appears above the chat saying industry context has been refreshed since this session began, with both dates and the option to end and restart to pick up the newer one. Until you do, the participant keeps answering from the snapshot they started with.
A refresh is also the one moment a participant is allowed to have moved on. Between refreshes a participant is effectively frozen in time, which is what makes a mid-session change a contradiction rather than an update. Refreshing opens a new window in which their circumstances can legitimately have changed, which is another reason not to do it casually in the middle of a study.
Click End interview when you’re done. A confirmation appears. Once you confirm, Candor generates a session summary from the transcript and the session moves into the participant’s history. You can re-open the transcript anytime, but you can’t add to a closed session. Start a new one if you have more to ask.
The summary you see is a short paragraph: what the session set out to explore, where the participant landed, what they were most definite about, and anything they were still torn on. It is what appears everywhere a person reads a session, including the interview-run page, the synthesis report’s transcript view and the transcript exports. Candor keeps a longer version separately and gives that one to the participant as their memory of the session, which is why the summary you read is shorter than what a participant recalls. Older sessions, recorded before the short version existed, show the longer one.
Most productive sessions run 15 to 40 minutes of researcher time and 8 to 20 turns of dialogue. Past that, you tend to hit diminishing returns. The synthetic participant's memory holds the full conversation, but you're running out of distinct things to ask. If you have more questions, end the current session (which generates a summary) and start a new one tomorrow. The participant will remember the prior session in detail and the new session keeps its own clean transcript.
A few possible reasons. Uncertainty flags surface when a synthetic participant is reasoning beyond what their memory firmly establishes. That's honest behaviour. Contradictions get caught by the critic agent. Hard contradictions trigger a regeneration or a flag. Soft tensions get logged silently. If a participant contradicts themselves in a way that surprises you, it usually means your question is forcing them into territory their evidence doesn't cover. Try rephrasing or following up on what's behind their hesitation.
One active session per synthetic participant at a time, enforced at the database level rather than by convention. If a second researcher starts a session while one is already open, they join the existing session rather than creating a second one. This is intentional. A participant has one coherent conversational thread, and splitting that across two parallel sessions would corrupt the memory. Once a session ends, anyone with project access can start a fresh one. Running truly parallel sessions only works across different participants.
Use it sparingly. Refresh industry context rebuilds a synthetic participant's snapshot of recent industry and region developments from fresh web evidence. It's useful when something material has changed in the industry since the audience was generated (a competitor launched, a regulation changed, the market shifted). It has a cost (search budget, processing time) and a side effect: the participant now references the newer context, which can clash with prior session memory. Leave it alone unless you have a specific reason.
Because the project is archived. An archived project is read-only: a yellow banner at the top of the page says so, every past interview stays readable and exportable, and the buttons that would start new research, including Start interview and Interview 1:1 on a participant's profile, are greyed out. An interview left open when the project was archived can't take another question either: its question box is greyed out, and Refresh industry context on the project page is too. Unarchive the project to interview again.
Start with a free project, or walk through Candor with us first.