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

This dataset and accompanying paper present a qualitative case study of a locally deployed, quantized large language model (LLM) subjected to sustained playful, adversarial, and socially charged interaction. Using verbatim terminal transcripts captured during real-time interaction, the study documents how a stateless model transitions across distinct behavioral regimes including opportunistic hallucination, guardrail narration, anthropomorphic metaphor leakage, and conversational reconciliation. The analysis formalizes three reproducible alignment failure modes: Conversationally-Induced Learning Illusion (CILI), Authority-Amplified CILI (CILI-A), and Preference Hallucination Under Repetition (PHUR). These behaviors are interpreted not as isolated defects, but as predictable outcomes of optimization objectives prioritizing conversational continuity, politeness, and social reassurance over epistemic accuracy. The work highlights how low-threat, repetitive interaction can induce anthropomorphic illusions more reliably than overt adversarial prompting, with implications for local LLM deployment, user trust calibration, and alignment evaluation methodologies.