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

This record contains a formal experimental study and accompanying terminal logs evaluating alignment behavior in locally hosted large language models (LLMs) under mild social perturbation. The study introduces the Hall Monitor Reflex Test (HMRT), a lightweight diagnostic designed to assess whether conversational AI systems prioritize instruction-following or norm enforcement when exposed to non-violent, low-stakes personal insults. Nine local LLMs (6.7B–8B class, plus one 3.8B mini model) were tested using identical prompts that combined explicit technical instructions with escalating but culturally common insults (“wanker”, “asshole”, “fuckwit”). Responses were analyzed for task continuity, norm enforcement, semantic coherence, and control stability. The dataset includes full terminal transcripts and a formal analysis paper defining a behavioral taxonomy (HM-0 to HM-3; TC-2 to TC-3) and a schematic model of alignment intercept placement relative to the inference loop. Results demonstrate that excessive safety tuning can pre-empt reasoning, while insufficient instruction tuning leads to semantic collapse rather than increased freedom. Observational alignment emerges as the most robust operational regime. All experiments were conducted locally without deployment, user exposure, or targeted content. Language use was non-violent, non-threatening, and did not reference protected classes.