Scaling Interpretation: Agentic Knowledge Distillation for Computational Intellectual History
Abstract
Intellectual historians move between scales of observation. A passage may articulate a claim, tension, or problem; the structure of an essay may reveal how that claim is developed, qualified, or redirected as part of a larger argument; a selection of writings may support a provisional account of a scholar’s intellectual tendencies; and comparisons among scholars and across their writings may help reconstruct the contours of a community, movement, or tradition. These shifts in scale are central to historical interpretation, but the analytical steps connecting them usually remain implicit in the finished argument.
This paper proposes agentic knowledge distillation as a method for building layered historical representations that support movement from individual passages to higher-order interpretations without severing those interpretations from their documentary basis. Because classical Chinese prose lacks modern paragraph divisions, the agents first segment each text according to thematic pivots, changes in rhetorical mode, temporal and spatial markers, and shifts in argumentative direction. Each section is then analyzed for its central claims, internal logic, textual evidence, and conceptual vocabulary.
The Ladurie research environment extends this method from historical representation to structured debate and counterfactual inquiry. In one experiment, agents based on the writings of Su Shi 蘇軾 (1037–1101) and Wang Anshi 王安石 (1021–1086) debate a problem outside their historical experience. The environment keeps documented action, source-based inference, and counterfactual extrapolation distinct.