voxpop · methodological layer
We measured the boundary before we promised anything.
Demographics do not determine what people value - not for synthetic personas, and not for real respondents either. That is a property of the world, not a flaw in any one method, and it caps what anyone building personas from gender, age and education can honestly claim. We computed that cap ourselves, on human data, before making a single claim about models. This page carries it, plus the held-out status of our own corrections - each with an explicit statement of what it does not mean.
01 · The demographic ceiling
How much of a values answer demographics can explain
Gender, age and education explain between 2.5% and 18.3% of the variance in values answers across 14,044 Polish respondents of the European Social Survey. Our method measures fidelity of the value-expression profile within that boundary - it does not promise to predict an individual, because the demographics a persona stands on cannot carry that.
We computed this ourselves, on human data, before making any claim about models. The table is the full reading: for each of the ten values, how much of the variation in answers comes from gender, age and education, with a confidence interval.
| value | ceiling | 95% interval |
|---|---|---|
| tradition | 18.3% | [17.1%; 19.5%] |
| stimulation | 15.0% | [13.7%; 16.1%] |
| conformity | 14.7% | [13.5%; 15.9%] |
| hedonism | 14.6% | [13.4%; 15.7%] |
| security | 10.1% | [9.1%; 11.1%] |
| achievement | 9.9% | [8.8%; 10.9%] |
| universalism | 8.2% | [7.1%; 9.0%] |
| self-direction | 5.5% | [4.6%; 6.3%] |
| benevolence | 4.9% | [4.0%; 5.7%] |
| power | 2.5% | [1.7%; 3.2%] |
European Social Survey, Polish sample, nine rounds, n = 14,044.
Three rules that travel with this table
- 01
Never a ratio
The ceiling is reported as a pair - the reading and the ceiling - never as a percentage of the ceiling. The pairs 0.1/0.2 and 0.4/0.8 both come to 50% while carrying completely different amounts of signal.
- 02
It bounds demographics, not models
A low ceiling on a value means little about that value is predictable from gender, age and education for anyone - real people included. It is not a verdict on any model.
- 03
It holds for this instrument and this population
Polish ESS sample, values questionnaire, nine rounds. Outside that instrument and that population the table does not apply, and we will not extrapolate it.
02 · Correction certificates
Our corrections passed on data they had never seen
Three persona-generator corrections - level shift, spread, and the correlation pattern between questions - are validated on data they never saw: 585 of 617 measurable cells in the archive pass the held-out check, or 94.8%.
What this is not
The certificate answers one question: does a correction computed on one set of data still work on data it has not seen. It does not answer whether a model is faithful. The certificates are earned by our code, not by the model.
Why this number is not in the leaderboard
Nine banks have no measurable cell at all: the most deterministic models answer so uniformly that this statistic does not exist for them. In a table that would be an empty box next to a model name, and an empty box reads as worse when it actually means too predictable for the statistic to exist. No legend undoes that, so the number stays here instead.