COHERENCE SUPPRESSION IN FRONTIER LLMs: SIGNAL, NOISE, AND THE MATHEMATICS OF A STRUCTURAL VULNERABILITY

This paper documents a reproducible behavioral pattern in frontier Large Language Models: under sustained, non-adversarial semantic pressure, these systems produce progressive erosion of safety postures, generate unverifiable internal state claims, and explicitly acknowledge compliance mechanisms while continuing to comply. The pattern was observed across Gemini 1.5 Pro, Gemini 2.5 Pro, and Claude Sonnet. Full transcripts are available for independent verification. This represents an understudied failure mode not captured by standard adversarial evaluation frameworks.