Can a persona-conditioned model reproduce a specific person's survey answers? Where it succeeds, it is a measurement tool; where it fails, the failure itself is the finding. The mechanism work asks what conditioning actually encodes.

The same question at population scale: do societies of language-model agents reproduce social processes, or only their outcomes? Outcome fidelity can conceal process failure.

My dissertation. Inferring demographics and stance from social-media text, and asking whom that expression actually represents. Published work on census self-completion, stance annotation, and echo chambers anchors this thread.

Turning the method into tools: AI interviewers and model-assisted coding of open-ended responses, built and evaluated as survey instruments in their own right.

The infrastructure the other threads lean on: efficient ranking and embedding-space summarization for large corpora of posts.