Towards Generative Relational Learning: A Preliminary Software Infrastructure Design

Learning-oriented software can make information easy to collect while leaving the roles, assumptions, histories, and possible uses of that information difficult to reconstruct. This paper develops a preliminary infrastructure for generative relational learning, understood here as a design orientation toward formulating further questions, explaining intellectual operations, inspecting relations, and judging possible transfers. The proposed infrastructure permits selected structure to emerge within ordinary writing. Independently identified questions, concepts, formulas, methods, and other authored entities can remain in a continuous note while becoming reusable through stable references. A bounded extension of Markdown distinguishes prose mentions from explicitly typed relations and retains the occurrences from which graph edges are derived. The implemented hosted prototype combines a Milkdown/ProseMirror editor, a source-preserving adapter, shared browser/server semantics, SQLite revision transactions, persistent numbering, and a separate proposal-review boundary. The paper reconstructs the design's development, distinguishes the hosted implementation from its earlier desktop proposal, and follows a synthetic volume-balance specimen through authoring, reuse, graph inspection, and revision. The supplied handoff reports 143 passing engineering tests; a separate re-execution for this paper passes 75 dependency-free parser and source-adapter tests. These results concern bounded software behavior. The paper develops a prospective evaluation programme for authoring burden, semantic comprehension, explanation, transfer, and substantive review, without reporting uncollected learner outcomes. Its contribution is an inspectable design and artifact account linking learning-oriented questions to concrete representation, interaction, and preservation contracts.