Establish the Global AI Safety Testing Authority (GASTA) with authority to conduct standardized pre-deployment evaluations of all AI systems above defined capability thresholds. GASTA operates 3 regional testing centers (Americas, Europe-Africa, Asia-Pacific).

Testing batteries cover: (1) catastrophic misuse uplift testing for CBRN and infrastructure attack capabilities; (2) societal impact stress testing at 100M, 500M, and 1B user scale; (3) robustness testing under distribution shift and adversarial conditions; (4) demographic fairness assessment across 40 protected characteristic combinations.

Labs submit models 90 days before planned launch. Results are public. Certification is required for market access in member states. Failed tests result in remediation requirements, not permanent prohibition — creating incentives to build safety in from the start rather than as a post-hoc add-on.

Modeled directly on automotive crash testing which transformed vehicle safety without requiring proprietary design disclosure.