Last week I was in Burlington, Vermont for IC2S2 — the International Conference on Computational Social Science, hosted this year by the Vermont Complex Systems Institute at the University of Vermont, July 28–31. I presented a poster from the political geometry project, reframed around the question the paper is really asking: what does the instrument identify? An audit of political representation in a language model is a measurement design before it is a political finding — the prompts, the readout site, and the choice of instrument each shape the structure you end up attributing to the model. Two instrument families, asked of the same 250 prompts across four open-weight models, do not identify the same thing.
One detail I keep thinking about: the program was assembled with embeddings — abstracts placed in a semantic space and clustered, so sessions group work that is genuinely close rather than work that happens to share a keyword. A conference on computational social science quietly running its own methods on itself, and it showed: the room I was in was full of people asking the same question from different disciplines, which is not what conference keywords usually get you.
Now: turning the audit argument into the revision I’ll present at APSA in Boston next month, moving the congressional speeches project toward a first draft, and getting the network paper out to a journal.