Treat evasion as part of the design
An agreement to prevent Artificial superintelligence: capability substantially beyond humans across a broad range of important tasks. would face a serious strategic problem: each participant could fear that others would cheat and gain an overwhelming advantage. A declaration of good intentions would not resolve that fear. Participants would need credible information about compliance, a response to violations, and confidence that following the rules would not leave them defenseless.
W360 proposes an international capability limit supported by verification. This is an institutional design goal, not an existing guarantee that every dangerous project could be detected. Its feasibility depends on how concentrated the relevant hardware and expertise remain, and how quickly prohibited capabilities could be developed in secret.
Build several overlapping checks
Possible measures include inventories of advanced compute, reporting by large compute providers, audited training records, inspections of declared facilities, and independent evaluations of deployed systems. No single measure establishes compliance. Hardware records might reveal capacity without revealing how it is used; model evaluations might miss concealed or untested behavior.
Inspection arrangements would need rules for protecting legitimate trade secrets and national security information. Technical assurances should be tested against deliberate evasion. An agreement should specify how inspectors handle inconsistencies, what access they can request, and how disputed findings are reviewed.
Make compliance a stable choice
Participants would need shared definitions, predictable consequences for violations, and practical benefits from remaining inside the agreement. Those benefits could include access to permitted research, trusted supply chains, and cooperative security arrangements. Responses to breaches should be lawful, proportionate, and collectively accountable.
Countries outside the agreement remain a problem. Coordinated export restrictions and conditions on access to participating markets could raise the cost of noncompliance, but they could also encourage rival supply chains. Policy design must examine both effects.
Define what success requires
Verification does not require perfect knowledge of every computer. It does require a strong enough chance of detecting consequential violations early enough to respond. If a clandestine project can cross the dangerous threshold before detection or enforcement matters, the agreement has a critical weakness.
A serious proposal should publish its assumptions about detection, development speed, response time, and rebuilding routes. Public confidence should rest on those assessed conditions, with regular independent challenge, rather than on the existence of a treaty alone.