Reasoning Scaffolds for AI-Assisted Behavioural Modelling: A Logical-Square Instantiation

Abstract

AI assistants are increasingly used to support behavioural modelling from natural-language requirements, yet they provide little assurance that the resulting models are logically coherent. We argue that this limitation stems not only from the capabilities of large language models but also from the absence of explicit reasoning structures mediating interactions between human analysts and AI systems. To address this limitation, we introduce the notion of a reasoning scaffold as an explicit intermediate modelling artefact and propose scaffolded behavioural modelling, in which AI agents reason over such scaffolds rather than rely solely on conversational interactions. As a first instantiation of this broader concept, we investigate logical squares as one possible reasoning scaffold that exposes contradictions, alternatives, implications and missing behavioural concepts, and introduce scaffold-aware agents that exploit these structures to support systematic behavioural exploration. Finally, we outline a research agenda for reasoning scaffolds and argue that they provide a foundation for more explainable and verifiable AI-assisted modelling environments, and may facilitate future correct-by-construction approaches.