Mitigating Sycophancy and Hallucination in Large Language Models: The LOGOS Case Study in the Era of Reasoning Models (DeepSeek-R1, Gemini 3.0, ChatGPT-5)

Reliability in large language models (LLMs) is constrained along two orthogonal axes: sycophancy—a social-alignment failure where the model tends to agree with user pressure—and hallucination (confabulation)—a world-alignment failure producing confident but unsupported content (Fanous et al., 2025; Sharma et al., 2023; Ji et al., 2022; Kalai et al., 2025). Building on this distinction, we introduce LOGOS, a domain-bounded Constitutional RAG system for Orthodox theology with explicit anti-sycophancy rules and strict scope boundaries. Prior to human comparison, LOGOS underwent a Phase1 theological audit: a committee of the Holy Metropolis of Nea Ionia posed 200 domain questions, finding all responses doctrinally sound. We then conducted a Phase2 human evaluation with N = 52 AI students (valid responses N=51– 52 per prompt), using five structured prompts plus one freeform redteaming task.