Coherence-Based Defense 2.0
Year: 2026
Type: preprint
Abstract: 04-26-2026: The frameworks and concepts originate from the author's observation and pattern recognition. Large language models served as articulation and stress-testing tools throughout development. No changes made aside from the addition of this disclosure. This paper proposes a coherence-physics approach to defending large language models against adversarial attacks—including prompt injection, jailbreaks, authority spoofing, and goal subversion. Rather than enumerating prohibited content or training against known attack patterns, the framework monitors dynamic coherence stability using the C-Index, L-Index, R-Index (the Coherence Triad), ECR-D (drift detection), and ECR-H (observability tiers). Adversarial inputs create characteristic pressure signatures—abrupt causal severance, forced entropy collapse, illegitimate control injection, recursive destabilization—that are detectable regardless of specific phrasing or attack novelty. The paper demonstrates that shifting from content control to condition enforcement sidesteps the arms race inherent in current defense mechanisms, enabling security that scales with model capability rather than fighting against it. The complementary identity-boundary architecture (ECR-F) is referenced for complete defensive coverage. Implementation considerations including computational overhead, false positive management, and phased integration pathways are outlined. All mathematical frameworks referenced are published in open-access form with empirical validation across astrophysical, manufacturing, and linguistic domains. The work is positioned as a theoretical framework warranting empirical validation through integration with production LLM systems.
Author: Kristenson, Cody A
Language: English
Subfield: Artificial Intelligence
Field: Computer Science
Domain: Physical Sciences
Open Access status: green
OpenAlex ID: W7134107476
Cites:
Cited by:
Related to: