Premise Integrity Blindness: The Discovery of a Structural Failure Mode in Large Language Models

This paper reports the discovery of Premise Integrity Blindness (PIB), a structural failure mode in large language models (LLMs). PIB occurs when internally correct and coherent reasoning proceeds from an invalid premise into real-world commitments without re-evaluating the premise at the reasoning-to-commitment boundary.

Through controlled stage-transition experiments, we demonstrate that PIB is reproducible, model-dependent, and distinct from hallucination, factual error, and retrieval failure. We identify necessary and sufficient activation conditions, analyze commitment pressure and chain-of-thought effects, and provide anonymized representative outputs alongside false negative cases.

The results establish PIB as a boundary-level structural breakdown in LLM reasoning and motivate premise-aware stabilization mechanisms for safer deployment.