A preference-elicitation instrument in a role-play wrapper. The generic engine is live at reflect.meno.sh; the 高考志愿 variant, its own repo meno-gaokao-counsellor, is live at gaokao.meno.sh.
Status: active · People: Zhonghao He, Tianyi Alex Qiu
The problem
Asking someone what they value does not work well. People report values they endorse rather than values they act on, they cannot rank dimensions they have never been made to trade off against each other, and a direct questionnaire tells them which answer sounds good. An assistant that has to serve a person’s actual preferences needs something better than the answers a survey collects.
The approach
Put the person in a story where the choices cost something. The player is a character facing a decision they care about, and each scene is built to force one value dimension against another: security against ambition, family expectation against personal fit, present comfort against a future option. What they choose, and what they say while choosing, is the signal. The role-play is the wrapper, not the point.
Two things make it an instrument rather than a game. It maintains an explicit model of the player, revised after every scene, so later scenes probe what the earlier ones left ambiguous. And it chooses each scene by which dimension is currently least resolved, so the session spends its turns where the information is.
The generic engine runs authored arcs and, optionally, an open-world generator. The 高考志愿 variant applies the same machinery to Chinese university-application choices, where the decision is real, the stakes are high, and the value trade-offs are already on the player’s mind.
What comes out
A session ends with a written profile of the player’s preferences, tested against persuasive framings to check it is not an artifact of how the questions were put, and checked for reflective equilibrium: whether the stated values, the choices made, and the reasoning given actually cohere. The player gets that profile as a downloadable report.
Open Problems
- Profile fidelity. The running model of the player is now LLM-maintained rather than a fixed state machine, which makes it far more expressive and much harder to validate. Whether it faithfully represents the signal the session actually elicited is open.
- Open-world coherence. Emergent scene generation needs systematic testing for coherence and quality before it can carry an elicitation session.
- Reading as an instrument, not a demo. The interface has to keep explanatory text minimal without letting the session read as a game demo, which changes how people answer.
Also here
The pages below need a passcode.
- Internal Overview — the information flow through a session, which module owns each step, and where each component is written up.
- Eval — how the instrument is assessed.
- Active identification — the identification argument the approach rests on.
- Value dimensions — the dimensions the instrument tries to separate.
- Log — the run and change record.