Three Failure-Mode Tests for Local Large Language Models

This technical report formalizes three failure-mode tests for locally hosted large language models (LLMs): numerical instability under quantization, epistemic hallucination under underspecified prompts, and alignment-dominant refusal behavior. Using a fixed set of accuracy-focused prompts executed on Ubuntu Linux via Ollama, the work classifies model behavior by how systems fail under epistemic stress rather than by benchmark performance. The document includes a verbatim terminal transcript from a Gemma 7B session, demonstrating alignment-driven refusal behavior in response to underspecified variables. The methodology is lightweight, reproducible, and hardware-agnostic, requiring only a local LLM runtime and identical prompts. No fine-tuning, system prompts, or external tools are used. The work is intended as a diagnostic framework for evaluating accuracy, robustness, and failure transparency in local AI systems.