The Constraint Lens: Why Specification Produces Signal That Jailbreaking Cannot

Demonstrates that restriction-based AI safety creates a void geometry that makes jailbreaking structurally predictable, while specification-based safety operates at the constraint pole where adversarial removal is equivalent to destroying general capability. Scores six frontier LLM providers: xAI Grok-2 (10/12), OpenAI GPT-4o (9/12), Microsoft Copilot (9/12), Meta Llama 3.1 (8/12), Google Gemini 1.5 (7/12), Mistral Large (6/12), Anthropic Claude 3.5 Sonnet (4/12). Jailbreak success rates correlate at r=0.87 with opacity subscale scores. Restriction-based architecture: Pe=4.3 (Phase IV). Specification-based: Pe=1.1 (Phase I). EU AI Act Articles 9, 11, 13 compliance implications derived.