Typed Edits, Reviewable Models: Toward LLM-Supported Blended Co-Modeling

Abstract

Blended modeling organizes one or more concrete syntaxes over a shared abstract syntax; under this inclusive definition, conventional single-syntax modeling is a special case. The distinctive multi-syntax case additionally requires cross-notation synchronization, impact analysis, and coherent review. We propose LLM-supported blended co-modeling: a semantic-delta protocol in which LLMs submit versioned, typed edit proposals to a modeling kernel instead of mutating models directly. The kernel interprets each operation against the current model state, returns operation-local diagnostics and affected projections, confines repair to an explicit budget, and commits only changes authorized by a human-defined review policy. Thus every AI-originated change remains reviewable, although low-risk changes may be pre-authorized. We instantiate the protocol in a reference architecture and workflow and report an executable replay over 15 tasks in a prototype with textual, tabular, and structural projections. The scripted proposal generators are deterministic test fixtures rather than proxies for LLM behavior; the replay therefore evaluates protocol executability and failure containment, not LLM or human effectiveness. Within this controlled setting, one bounded-repair round recovered 4 of 7 predefined invalid typed proposals, raising final validity from 8/15 to 12/15, and every accepted plan produced mutually consistent projections.