Domesticating the Latent: How Topic Models Become Evidence across the Social Sciences

Topic models travel across the social sciences as recognizably common formal techniques, yet their outputs do not arrive as ready-made substantive findings. Researchers must decide which outputs to retain, how to interpret and reorganize them, what analytical work they should perform, and which resulting objects require assessment. This study examines how these decisions transform formal model outputs into credible evidence. We conceptualize topic-model use as a representational career and combine comparative close reading of 32 purposively selected studies with a corpus-level analysis of reported operations in 1,017 empirical articles published in sociology, communication, and political science journals. The analysis identifies no single dominant passage from fitted model to published finding. Reported decisions are commonly oriented toward the intended claim, target construct, or requirements of a receiving method, while non-focal topics may remain numerically operative after disappearing from the substantive narrative. Topic-derived objects also differ in how and where they depend on stabilized meaning, ranging from semantically constituted measures to label-independent calculations and infrastructural uses in retrieval or qualitative analysis. Reported assessment events are concentrated on upstream model and topic objects. Comparative close reading further shows that when assessed outputs are subsequently transformed, a new assessment does not consistently follow. Disciplinary differences are patterned but distributional: journal fields alter the relative prevalence of particular op- erations without determining the career of an individual study. We introduce the concept of a completion regime to describe how published research workflows distribute interpretive labor, representational closure, and public warrant across the chain from algorithmic output to substantive claim.