The Lumen framework for generating structurally adapted user interfaces under design-system constraints

Replication package for the study "The Lumen framework for generating structurally adapted user interfaces under design-system constraints". Lumen is a framework that uses a Large Language Model to generate user interface specifications adapted to predefined user profiles, including users with accessibility needs, while restricting the generated interfaces to the components of an established design system. The package contains the material needed to understand, assess, and replicate the four phases of the research: systematic literature review, framework design, experiment planning, and data analysis. 01-RSL Details (systematic literature review): review protocol (research questions, search string, databases, selection criteria, and quality assessment); raw search results from the five databases in BibTeX; all retrieved articles with their final status and the exclusion criterion applied; and the data extraction form of the 11 primary studies. 02-Framework Details (framework design): proto-persona definitions and user journeys; system requirements specification; detailed description of each framework component by architectural layer; and the hypotheses and diagnostic metrics proposed for future end-user evaluation. 03-Experiment Planning (test cases and generated interfaces): definition of the 8 test cases (4 proto-personas x 2 tasks); case data supplied to the generation pipeline; the four proto-persona profiles; the eight generated interface specifications in JSON; and 66 screenshots of the evaluation application and of the generated interfaces rendered in a browser. 04-Data Analysis (technical and expert evaluation): the allowed component vocabulary (44 component types) and the Design System Adherence Rate of each interface; anonymized profiles of the 20 expert evaluators; the 160 raw interface-level evaluations; aggregate results by criterion, interface, proto-persona, and session position; structural metrics and generation time of each specification; provider-reported token usage; a statistical note describing how each statistic was computed; and a Python script that recomputes every aggregate table and statistic from the raw files. Institution names, personal names, addresses, and service endpoints appear as generic placeholders in the data files. Interface identifiers TC-01 to TC-08 are consistent across all files. License: Creative Commons Attribution 4.0 International (CC BY 4.0).