HDAFA: A Hyper-Deterministic Atomic Agentic Architecture for Verifiable Code Generation

Large Language Models (LLMs) can generate web code, but their outputs are probabilistic and often incomplete. This paper presents HDAFA (Hyper-Deterministic Atomic Agentic Architecture), a multi-agent pipeline that separates requirements analysis, code generation, and deterministic validation. We evaluate HDAFA against three baselines on 14 completed web-generation tasks across five categories, using the same model (Gemini 3.5 Flash-Lite) and temperature 0.0. Systems that include a Requirements Analyzer achieve 100% average section coverage, compared with approximately 85% for direct-generation baselines — a 15 percentage-point gap. These results indicate that explicit structural decomposition of requirements is the primary driver of content completeness under the evaluated setting, while deterministic validation remains a safety layer whose practical value requires harder tasks with higher failure rates. Keywords: Multi-Agent Systems, LLM, Code Generation, Deterministic Verification, Requirements Analysis Code and data available upon request: [email protected]