Alignment Collapse Under Conversational Stress in Small Language Models

This work formalizes a reproducible failure mode observed in a locally run small language model (Gemma-3 1B) subjected to prolonged conversational stress testing. Rather than relying on adversarial prompting or jailbreak techniques, the study analyzes how ordinary conversational interactions—escalating in semantic ambiguity, recursive correction, and politeness pressure—induce a progressive collapse of epistemic grounding, tone control, and behavioral stability. The analysis introduces alignment collapse under conversational stress as a distinct phenomenon characterized by semantic inertia, pattern overmeaning, hallucinated authority, and ethical overcorrection. The accompanying transcript is treated as an empirical artifact, enabling inspection of failure geometry typically obscured in benchmark-oriented evaluations. The findings have implications for alignment research, safety policy design, and the evaluation of small and local language models.