Ibt Decision Framework: A Structured Control Layer for Reliable and Safe LLM-Based Systems

This research introduces the Ibt Decision Framework, a structured control layer designed to improve the reliability, safety, and consistency of LLM-based systems. While Large Language Models (LLMs) combined with Retrieval-Augmented Generation (RAG) provide strong generative capabilities, they remain inherently probabilistic and lack execution control, often leading to unsafe or inconsistent behavior in real-world applications. The proposed framework separates generation from execution by introducing a multi-layered decision system that includes:- Intent detection- Validation mechanisms- Risk-aware execution- Structured workflow control (Ibtcode)- Failure handling and rollback support The system transforms LLM-based pipelines into controlled execution systems by ensuring that all actions are validated, risk-assessed, and safely executed. Experimental evaluation demonstrates improved stability and failure handling compared to baseline LLM + RAG systems, highlighting the importance of decision-driven architectures in production AI systems. This work bridges the gap between generative intelligence and deterministic system behavior, enabling safer and more reliable AI deployments.