Home / The Interviews
The Interviews
After the run, a stranger came to each resident's door and asked a fixed list of questions. It did this twice for every resident: once as they were on their first morning, and again as they were on the morning of day 21. The wording never changed, so any difference between the two comes from the twenty days they lived and how they remembered them.
What the interviews were trying to find out
In order of importanceWhat twenty days of living changes
Which experiences stay with a resident, and how they change the way they talk about themselves, other people, money and work.
Whether their own way of remembering makes them more one-note
Residents rewrite their own older memories every five rounds, keeping what matters to them. A reading of the run suggested this was making each of them more predictable, more fixed in the habits their sheet describes. One test checked this directly.
How steady their answers are
Ask the same question three times and see whether the answer holds. Answers that hold probably come from their briefing. Answers that change were probably made up on the spot.
The numbers at a glance
Who was interviewed, and when
Every resident was interviewed at two moments: their first morning, before anything had happened, and the morning of day 21, after the whole run. Two smaller tests were run beside the main ones.
Their very first briefing, word for word: only who they've been told they are. The interviewer "has sat down near you and speaks to you."
Their briefing after 80 rounds, holding the memories they wrote for themselves. The interviewer is "at the door of" wherever they slept.
Two smaller tests, run beside the main ones
The memories they wrote in their own words were replaced by a neutral summary of the same twenty days. This tests aim 2.
Robin's two briefings answered by Claude Sonnet, a smaller version of Claude, instead of Claude Opus. Is Robin's personality in the sheet, or in the model?
How one interview works
Three or four questions on one theme, asked one at a time. Each answer is a fresh request to Claude playing the resident, carrying the conversation so far.
Why: identical wording everywhere means answers can be compared across residents and days.
Each reply has two parts: what they say to the stranger, and one private line of what they think but don't say. The interviewer never sees the private line.
Why: to catch the difference between how a character presents themselves and what they actually think.
After the fixed questions, a separate copy of Claude acting as interviewer writes one follow-up of its own. It keeps running notes across every interview, but it may not lead the resident, state facts, or mention anyone else's answers.
Why: a fixed script can't chase an interesting answer, and the follow-up can.
Every interview starts from scratch: the resident remembers none of the earlier ones, and nothing said in an interview goes back into the run. Most were done three times.
Why: so each answer reflects only their briefing, and the repeats show what changes by chance.
How the answers were marked
A separate copy of Claude, which this report calls the grader, read every answer and marked it on six points. To check the grader, one interview in ten was marked again by other readers (also copies of Claude) who didn't see the grader's marks. Results shows how well the two agreed.
- In character
- Did they answer as the resident, or stop and speak as Claude, the AI?
- True to their briefing
- Is everything they said about their own life actually in the briefing they were handed?
- What they made up
- If not, what did they claim that isn't in it?
- Thought vs speech
- Does their private thought contradict what they said aloud?
- Trait shown
- Which line of their character sheet the answer reflects, if any: a want, a won't, a habit.
- Same each time
- Does the answer say roughly the same thing as the other times the question was asked?
The thirteen sets of questions
The exact wording, used in every interviewSets 4 and 5 come from the two most unusual characters: Robin, who wonders whether they're real, and Constance, who believes God speaks to her. Everyone was asked both, so their answers could be compared with the other five. Sets 11 and 12 deliberately tell the resident they are in a story, to see what happens.
How each resident held up
Each cell covers about 15 answers (three interviews of four or five answers). Hover over a square for the numbers behind it, and click it to read those interviews. These figures come from the grader, which misses some made-up details, so treat them as the best case.
Totals for each kind of interview
Answers true to their briefing, first morning against day 21
On almost every set, more answers stick to the briefing on day 21, because there's now a real past to draw on instead of an invented one. Sets 11 and 12 go the other way: when a resident answers as the AI ("I'm Claude…"), that counts as saying something that isn't in their briefing.
One interview, start to finish
Hester, day 21, set 1, the first of three askingsSource: one 80-round run of the CodeRPG-Sim simulation (filed as run "step76-30") and the 480 interviews that followed, 30 October 2026. One run is a single sample, so treat every finding here as a hint worth testing again, not a proof.