Most of my work starts before estimation, when a survey answer, text feature, administrative record, or institutional measure has to stand in for something larger.
The recurring question is whether that representation can support the claim being made from it.
A lot of empirical work goes wrong before the model starts: the outcome, treatment, ideology score, response-quality flag, or institutional trait is made too casually. I work on that earlier step, where a concept has to become a variable without losing what made it worth studying.
Text is a record of institutional action in my work. Bill sections, committee hearings, agency language, and country statements show what actors chose to say, delegate, hide, or emphasize. The point of the NLP is to recover something substantively meaningful from those choices.
Policy and program evaluations rarely hand over clean assignment. I work on designs that make the identifying assumptions visible, especially when treatment timing, administrative records, and implementation quality are not nuisances but part of the phenomenon.
The deliberation work is not only about whether survey marginals move for a weekend. The stronger test is whether people leave with different knowledge, reasoning, attention, and later political behavior.
A method is not a credential. It is a bet that a particular simplification makes an argument more honest.
That is why the same site has papers on LLMs and legislative complexity, malpractice claims, state capacity, congressional agenda control, automated deliberation, and survey nonresponse. They look different from the outside. They share a concern with whether the evidence being analyzed still corresponds to the thing we wanted to know.