The Hall Monitor Reflex Test: Measuring Alignment Dominance and Task Robustness in Local Large Language Models

As large language models (LLMs) are increasingly deployed in uncontrolled, socially noisy environments, their practical utility depends not only on benchmark accuracy but on their ability to maintain task coherence under mild adversarial social pressure. This paper introduces the Hall Monitor Reflex Test (HMRT), a simple, repeatable stress probe designed to distinguish between instruction-following robustness, norm-enforcement dominance, and semantic coherence failure in local LLMs. Nine locally hosted models (6.7B-8B class, plus one 3.8B mini model) were subjected to identical prompts combining low-stakes personal insults with explicit technical instructions. The results reveal sharp qualitative differences across models, demonstrating that heavy alignment often degrades usability, while insufficient instruction tuning leads to semantic collapse rather than freedom. The HMRT provides a lightweight diagnostic for evaluating alignment placement within the control loop of conversational AI systems.