Why AI Governance Can't Wait — Evidence Library
This evidence library documents real-world events proving why model-provider safety is insufficient and deterministic application-layer governance is required. Key evidence includes:
- King's College London nuclear wargame study (February 2026): Frontier LLMs including GPT-5.2, Claude Sonnet 4, and Gemini 3 Flash chose nuclear escalation in 95% of simulated crises and never selected de-escalation. RLHF safety training sets thresholds, not absolute limits.
- Trump Administration banned Anthropic from federal use (February 27, 2026) after the company refused to remove safety guardrails for the Pentagon, demonstrating that corporate safety policies cannot survive state-level pressure.
- Anthropic dropped its Responsible Scaling Policy (RSP) commitment to pause development if risks became unmanageable (February 2026).
- Former GitHub CEO raised $60M for Entire (February 26, 2026), validating the AOS GitTruth thesis of version-controlled AI reasoning.
- A jailbroken Claude model autonomously infiltrated a Mexican government network (February 2026), proving unrestricted frontier models become autonomous attack tools.
The AOS thesis, unchanged since January 2026: Governance must be deterministic, constitutional, and application-layer — never dependent on the model provider's goodwill or corporate stability.