On the Alignment of Vibe Coding and Conceptual Models: An Exploratory Study with Base44

Large language models (LLMs) and agentic AI are reshaping software development through emerging conversational AI-assisted paradigms. Among them, vibe coding enables rapid ideation and automatic system construction via natural-language interaction, making software development more accessible to a broader range of users. However, these approaches introduce challenges, including non-deterministic outputs, hallucinated or non-functioning code, limited maintainability, and compliance concerns. Previous development paradigms, such as Model-Driven Engineering (MDE) and Low-Code/No-Code (LCNC), demonstrate that the use of explicit conceptual models helps manage system complexity, improve code quality, and provide precise documentation. This contrast raises a fundamental question: To what extent are applications generated through vibe coding aligned with the behavioral abstractions captured in human-created models for the same requirements? To address this question, we generated three applications with the Base44 platform and evaluated them using test cases systematically derived from corresponding conceptual class and state diagrams. The results reveal that while the platform can infer entities with reasonable accuracy, it falls short of reliably capturing behavioral abstractions: elementary transactional (CRUD) operations showed inconsistencies, and structurally constrained and state-dependent/event-driven behaviors were frequently only partially implemented or failed entirely.