Guardrail Bloom and Hallucination Under Adversarial Play: A Case Study of a Quantized Local Large Language Model

This paper presents a qualitative case study examining the behavior of a locally deployed, quantized large language model (LLM) under progressively adversarial, playful, and socially charged interaction. Using a verbatim terminal transcript captured from a consumer-grade workstation, we analyze how the model transitions across distinct behavioral regimes: task compliance, opportunistic hallucination, template-based humor generation, de-escalatory guardrail activation, anthropomorphic metaphor leakage, and conversational reconciliation. The exchange provides a compact but revealing probe of alignment mechanisms, failure modes under null or ill-posed tasks, and the limits of humor and selfreferential language generation in small, quantized models. Rather than treating these behaviors as isolated defects, we frame them as predictable outcomes of optimization objectives prioritizing conversational continuity, safety compliance, and tone management over epistemic rigor.