Co-developing verification for Pacing the Frontier
Background
Frontier AI labs may eventually need to pace or constrain development or deployment at a critical capability threshold. Doing that safely, and giving other labs or governments confidence that commitments are actually being kept, depends on verification mechanisms that exist before they are urgently needed.
Without them, labs and policymakers could end up facing high-stakes decisions with very little room to move. That raises the risk of losing control of advanced systems, or of losing the ability to shape decisions about them (for example through blunt government intervention).
"Pacing the frontier" defines the problem but not the solutions. The statement asks government to enable optionality. We tentatively think this is backwards: frontier labs are more likely than governments to have the expertise and resources to get verification tooling working before it's needed.
Meanwhile, the verification community is building technical designs, experiments, and prototypes. What's missing is input from people inside frontier labs on whether any of it could actually work in a frontier environment. And the biggest open questions are ones only labs can answer: when would a mechanism need to be deployable, does it fit existing infrastructure, does it fit the lab's security posture, and what other operational constraints would make it impractical?
Kicking off the co-development series
This is the first event in a co-design series on moving verification from designs to deployment. The idea is to progressively narrow a large pile of verification proposals down to a small number of credible candidates:
(a) Agree with labs on which claims actually matter for pacing, at different time horizons.
(b) Paper-level red-teaming with lab infrastructure, security, and other relevant teams, to catch fatal constraints before anyone sinks effort into prototypes.
(c) Proof-of-concept builds, initially outside labs and on lab infrastructure where that helps.
(d) Technical and operational red-teaming of whatever survives-
This first session is stage (a). We know people at labs don't have much bandwidth to review proposals, so we won't ask you to evaluate a stack of submissions.
What we will do in this session
Ahead of the session we'll circulate a draft list of claims that labs or governments might want to make credible to each other. It's a starting point only. Please challenge it, refine it, reprioritize it, and add to it. We especially want people who think the whole verification agenda is misguided to tell us why.
Together, we will:
Identify and prioritize claims across training, inference, and the clusters they run on, focusing on claims that would matter at the point they need to matter, not just ones that are technically interesting.
Rank who those claims would be valuable for: lab-to-lab, lab-to-government, and interstate contexts such as the United States and China.
Sketch possible technical and operational mechanisms for verifying the highest-priority claims, including their assumptions, limitations, and realistic timelines to deployment.
Start mapping the path from design to deployment: what needs testing, what access or cooperation is required, who needs to be involved, and under what conditions. The goal is to iterate toward a pilot proposal that labs could take and run with.
Format is a structured brainstorm, small-group breakouts, and a closing synthesis. There will be food. We want this to be a good hang as much as a working session.
Everyone participates in a personal capacity. Participation doesn't imply endorsement by your employer. The discussion is under the Chatham House Rule.