Making LLMs Reliable When It Matters Most: A Five-Layer Architecture for High-Stakes Decisions