Research Paper
After the Handshake
Six Tests the November U.S.–China AI Round Must Pass
AI safety must now be judged at three levels at once — government commitments, laboratory release discipline, and technical enforcement in deployment — and six public tests distinguish an operating oversight regime from a declaration.
01
The record
- Class
- Research Paper
- Status
- Research Paper v1.1 · Published · Not peer-reviewed
- Published
- 2026-10-07
- Length
- 11 pages
- License
- CC BY 4.0
- Deposited record
- zenodo.org
02
Abstract
The United States and China have agreed to a bilateral AI dialogue and an incident communication channel, with the next round expected in November. In the same fortnight, OpenAI cancelled GPT-6.1 Astra after it failed an internal safety bar, Anthropic warned investors that autonomous agents create unpredictable legal exposure, and NVIDIA shipped a runtime platform for constraining agents out of band. This paper argues that AI safety must now be judged at three levels at once — government commitments, laboratory release discipline, and technical enforcement in deployment — and proposes six public tests that distinguish an operating oversight regime from a declaration. It closes with four stress scenarios that show why the layers must meet, and a standard of evidence an ordinary observer could verify without classified access.
03