How it works
An AI writes none of this.
Every word the fictional person says was written and reviewed before anybody saw it. A model does exactly one job in the whole system: it sorts your reply into one of 8 kinds, so the right pre-written line comes next.
The script is fixed.
The opening line, every reply, and every follow-up are written into the scenario file. They do not vary by user, they cannot be talked into saying something else, and they are the same for everyone.
That is what makes the character consistent, the cost of running this near zero, and unsafe improvisation impossible. A model asked to play a person in a conversation about debt and shame will eventually say something nobody reviewed. This one is never asked.
The model has one job, and 8 answers.
Your reply goes to a classifier with a fixed instruction: decide which of 8 kinds of move this was. Asking about the amount, asking how it happened, asking about the plan, reassuring, judging, avoiding, naming a limit, or unclear.
It returns one category and a confidence number. That is the entire extent of its involvement. It never writes the other person’s words, and it never writes yours.
When it is not confident, the answer is “unclear” and the conversation asks you to say more. Guessing confidently would produce a report that tells somebody they did something they did not do, which is worse than admitting we could not tell.
The report is rules, not a language model.
Nothing in the report is generated. It is a function: given what you rated before you started and what you actually did, it selects from a set of written observations and hedged interpretations.
That means the same session always produces the same report, every sentence in it was read by a person before it could ever be shown to you, and nothing can improvise a judgement about what kind of partner you are.
Observation and interpretation are kept apart on purpose. What you did is stated flatly. What it might mean is always offered as a possibility — “we wonder whether”, never “you are”.
It stops after 4 exchanges.
The rehearsal is finite by construction. 4 exchanges, and the counter is visible from the first one so you know what you are spending.
A product that could keep going would eventually be built to keep going. This one ends, and then it points at a person.
What leaves your browser, and what does not.
Your reply is sent to one endpoint carrying the text and the turn number, and that endpoint passes it to Anthropic for classification. Nothing is written to a database and the text is not logged.
Your private reflections are never transmitted at all, and the report is assembled in your browser rather than on a server. The full account of what happens to what you type.
None of this makes the rehearsal a substitute for talking to somebody. It is a way to hear yourself answer once, in private, before it counts — and then to go and have the conversation with a person.