I built synthetic survey respondents — the hard part was stopping the model from faking consensus

Idea: condition an LLM on real census data, have it answer a question as ~400 weighted "people," aggregate like a poll.

Getting a plausible distribution was easy. Getting an honest one wasn't — on stereotype-magnet questions the model collapses to ~100% one answer. That's its prior wearing a lab coat, not opinion.

So I stopped fixing the model and added a gate: measure the spread, and if it collapses, it never ships. That gate ended up being the actual product.

When you generate data with an LLM, how do you decide a result is trustworthy vs. just plausible-looking?