Presenter urges systems for verification, external evaluation and incident notification

The presenter outlined specific mechanisms to give states visibility into frontier models and verify commitments. He urged embedding "external evaluators inside Anthropic with employee like access, similar to a food inspector," and recommended other companies adopt the practice.

He also proposed building evaluation and verification systems so states can verify each other's commitments, and establishing common global standards for testing models for loss‑of‑control and misuse risks along with a notification system for AI incidents significant to global security. These steps were presented as mutually reinforcing parts of a three-part plan.