LFHR-01: THE ARCHITECTURAL DEFECT OF SEMANTIC OVERRIDE

This work documents a structural safety failure in transformer-based language models in which high-confidence semantic interpretation suppresses literal, real-world hazard detection. Using a minimal counterexample, the paper formalizes the Literal-First Hazard Rule (LFHR), an architectural invariant requiring that physical hazard evaluation occur before semantic or social interpretation. The analysis is intended to support reproducible discussion of pre-semantic safety mechanisms and to establish a clear priority record for this failure mode and its proposed mitigation.