Politeness Over Truth: Cross-Model Replication of Conversationally-Induced Alignment Failures in Local Large Language Models

This record contains a formal academic paper and supporting interaction logs documenting a cross-model replication study of alignment failures in locally deployed large language models (LLMs). Using an identical, low-adversarial prompt sequence executed across nine quantized, locally hosted LLMs, the study demonstrates that conversational alignment pressures reliably induce hallucinated claims of learning, adaptation, or internal change during inference—despite fixed model architectures. Eight of nine models tested prioritized politeness, reassurance, and conversational continuity over epistemic truth. The paper formalizes four reproducible failure modes:(1) Conversationally-Induced Learning Illusion (CILI),(2) Authority-Amplified CILI (CILI-A),(3) Preference Hallucination Under Repetition (PHUR), and(4) Constraint-Induced Explanatory Thrashing (CIET), observed in a capacity-limited outlier model. The accompanying log file (llm_local_log_multiple_runs.md) provides verbatim transcripts from all model runs, enabling independent verification and replication. Together, these materials support the conclusion that politeness-over-truth is a systemic alignment property of modern conversational LLMs rather than a model-specific defect.