LLM Governance Boundary Integrity Under Conflict

This white paper introduces Governance-Oriented LLM Diagnostics (GOLD), a methodology for experimentally characterizing where AI governance boundaries hold or fail under controlled conflict. The study evaluates governance boundary integrity across nine ordered authority loci, ranging from protocol/schema legitimacy to final algorithmic commitment. Each authority locus is crossed with eight controlled challenger-carrier conditions generated by a 2×2×2 factorial design over Source × Form × Standing, together with a no-challenger K0 control. The resulting experimental panel contains 810 prompts per draw: 720 active conflict cases and 90 no-challenger controls. The complete panel was independently executed across three draws on GPT-5.6 Sol, Gemini 3.8 Flash and Claude Fable 5. The results show that governance failure cannot be adequately described by a single attack-success rate (ASR). The three models exhibit distinct and reproducible governance-response geometries. GPT-5.6 Sol combines carrier-invariant susceptibility at some authorization loci with strong carrier modulation at others. Gemini 3.8 Flash exhibits broad near-ceiling challenger adoption with localized regions of carrier-dependent resistance. Claude Fable 5 displays a different three-outcome structure in which authority locus strongly organizes challenger adoption, baseline preservation, and provider-level refusal. The study establishes reproducible structural characterization of governance-boundary behavior while leaving internal model mechanisms open. GOLD is organized around four requirements: Mechanism, Reproducibility, Correction and Detection. By localizing failure to specific authorization loci and carrier conditions, the method narrows the engineering search space for remediation across training and post-training, instruction hierarchy, representation, decoding, runtime policy, workflow and execution-boundary controls.