DAVINCI-Service Boundaries and Interaction Recovery in Agentic Microservice Design: A Configuration Study

This deposit contains the complete reproducibility package for a controlled study on microservice architecture generation using multi-agent LLM pipelines. The corpus comprises 360 observations resulting from the full factorial of 3 language models (Gemini, DeepSeek, Claude) × 5 pipeline configurations (C0–C4) × 8 subject systems (7ep, AcmeAir, Cargo Tracker, DayTrader7, Jokul, JPetStore, PetClinic, TNTConcept) × 3 independent rounds, executed between 02 September and 24 September 2026. For each execution, the deposit provides three artifacts: the metrics PDF (containing service and interaction F1 scores against human-authored reference architectures, plus the validated YAML specification), the architecture diagram PDF, and the execution metadata JSON (tokens, wall-clock time, LLM calls). Alongside the outputs, the deposit includes all inputs (system requirements, reference services and interactions, static-analysis evidence, and ground-truth architectures), the pipeline (agent definitions, prompts, orchestrator, and evaluator), two statistical analysis pipelines (frozen and extended, both with fixed seeds), the consolidated report, and the qualitative human evaluation (protocol and anonymized responses from 12 evaluators). The package is organized by role rather than by checklist section, and a ZENODO-DEPOSIT-TREE.md file maps each requested artifact to its actual location. All primary evidence needed to reproduce every number in the accompanying paper is present; the six declared gaps (S4 robustness analyses, anonymized stimuli, respondent manifest, consent note, SHA-256 checksums, and deposit README in S1–S7 format) are documented as pending work rather than missing evidence. Keywords: large language models, multi-agent systems, microservice architecture, software architecture generation, empirical software engineering, reproducibility, ablation study, agentic AI License: CC BY 4.0 (license PDF included; final license posture to be confirmed before upload).