Induced Meta-Reasoning Failure in a Local Large Language Model: A Case Study in Instruction-Hierarchy Collapse

This work documents a reproducible failure mode observed in a locally hosted large language model (gpt-oss:latest via Ollama) triggered through a structured escalation of meta-reasoning demands. Unlike common adversarial or stress tests, the failure was induced without malformed input, prohibited content, excessive token length, or resource exhaustion. The crash emerged from a conversational instruction paradox in which the model was simultaneously required to respond and explicitly instructed to produce no output. The resulting runtime failure illustrates a semantic—not computational—breakdown in instruction hierarchy resolution, herein termed Semantic Stack Exhaustion. The accompanying materials include a formal academic report, full conversational transcript, and system context sufficient for independent replication. This artifact contributes to research on LLM robustness, instruction-following limits, and non-hostile adversarial prompt design, with particular relevance to locally deployed inference systems lacking graceful failure handling.