A Boundary-Aware Synthesis of Llm Extraction Defense And Physics-Editable World Models: Robust Evaluation under Distribution Shift

Abstract

A central challenge in LLM extraction defense and physics-editable world models is to compare studies whose mechanisms and validation settings do not share a single denominator. The present review uses a honeypot knowledge graph that redirects model-extraction queries toward low-transferability knowledge and a large-scale dataset organized around physically editable world-model factors as focal cases for a boundary-aware synthesis. Two target papers are triangulated against 12 locally validated publications. The comparison follows attack modeling, honeypot knowledge, query economics, utility preservation, adaptive attackers and deliberately separates mechanistic interpretation from performance ranking, because the latter can conceal incompatible experimental or operational conditions. Across the evidence base, the decisive issue is alignment: attack modeling shapes what is observed, honeypot knowledge shapes how it is compared, and adaptive attackers governs whether the conclusion can be transferred. Uncertainty is most informative when reported as part of the result rather than treated as a postscript. The resulting framework supports reproducible comparison while preserving differences between study designs, and it identifies concrete points at which transfer claims should be narrowed or retested.