Position: Assumption Search, Not Proof Search: AI Assistance for Theory-Oriented Social Science

This position paper argues that AI-for-math paradigms do not transfer directly to theory-oriented social science. Although theoretical social science—especially areas such as theoretical economics—often contains substantial mathematical derivation and proof, its central research bottleneck is frequently not proof search under strong verification. Instead, progress often depends on assumption search: identifying assumptions, mechanisms, boundary conditions, and interpretations that are coherent, meaningful, and disciplinarily acceptable under weak external validation. This difference has important implications for AI system design. We argue that AI assistance in such settings should be organized around human-in-the loop workflows that support problem framing, assumption proposal, verification, branching, and scholarly adjudication, rather than around proof-centric automation alone. We use Adrasteia as a concrete instantiation of this position and present limited empirical evidence as supporting illustration rather than definitive validation. Our goal is not to claim that theory-oriented social science can be reduced to workflow engineering, but to argue that future AI systems for these domains should treat human judgment and assumption search as first-class design principles. The paper is centered on economics and offers only cautious extrapolation to adjacent theory-oriented social-science settings.