Closed-Loop Optimization Risks: The "Fluent Hallucination" Paradox in Large Language Models

Closed-loop recursive generation has been hypothesized to induce instability in large language models (LLMs), typically manifesting as structural collapse in smaller architectures. However, the behavior of large-scale frontier models under sustained high-entropy recursion remains insufficiently characterized. We evaluate semantic and structural stability in GPT-5 Standard under 50-step closed-loop generation at temperature T=1.0. Results reveal a "Fluent Hallucination" phenomenon: the model maintains high lexical diversity (TTR ≈ 0.41) and extended output length, yet suffers from catastrophic semantic drift (mean > 0.90) across all domains. Crucially, run-level analysis (N=23) refutes the hypothesis that trajectory length protects against drift (Spearman ρ = +0.38, p=0.07). We conclude that model scale acts as a semantic mask: it preserves structural coherence while failing to anchor semantic logic.