Domain Knowledge Debt: Why Metamodeling Still Needs Elicitation

Abstract

We teach metamodeling because building a metamodel forces a student to understand a domain: its concepts, classifications, relations, and constraints are acts of abstraction, and abstraction is where domain understanding forms. The models, transformations, code generators, and modeling environments built on a metamodel all presuppose that understanding, so any deficiency propagates to every artifact derived from it. Generative AI is by now an accepted aid across software activities, and metamodeling is no exception: a large language model (LLM) can return a valid, often plausible metamodel in a single shot, so the artifact appears while the decisions that would have produced it do not. The metamodel can be correct, yet the modeler may lack the conceptual competence it presupposes. We call this gap domain knowledge debt, a debt of the learner rather than the artifact, invisible at delivery and paid later, when the metamodel must be extended, justified, or transferred. Building a metamodel step by step is itself knowledge elicitation that grows the modeler’s understanding, assimilating and accommodating cognitive schemas in Piaget’s sense, whereas one-shot generation extracts the same domain knowledge at once without the modeler acquiring it. Our position is pedagogical: however the LLM enters the workflow, teaching must keep the abstraction decisions with the student, preserving the incremental elicitation a normal modeling process induces rather than avoiding the tool. We instantiate this as Socratic, step-by-step construction, a scaffold that elicits each abstraction decision rather than an oracle that supplies it, illustrated through a five-phase sequence. This is a position paper; an empirical study is future work.