Most proposals for agent identity assume the same tradeoff: to make an agent accountable, you have to make it identifiable to whoever it's dealing with. A paper describing a national infrastructure project in China takes a different path. The system is described as national infrastructure and is scheduled for public launch in Q3 2026. The core design choice is a split. An agent is bound to a verified legal principal — a real person or organization who can be held responsible for what the agent does — but that binding is not disclosed to the business-layer participants the agent interacts with. A merchant, platform, or counterparty agent sees that it's dealing with an accountable entity, without learning who that entity actually is. The paper frames this as an underexplored point in the design space, distinct from the default assumption that accountability requires identifiability. The mechanism is called split-knowledge binding: no single party holds the complete mapping between agent and principal. Instead the knowledge needed to unmask an agent is divided, presumably so that resolving identity requires cooperation across separate holders rather than being available to any one party — a regulator, a platform, or the agent's own operator — acting alone. That's a meaningfully different trust model than the 'always show your ID' approach most Western identity proposals default to, where the agent's credential or DID is verifiable directly by whoever it talks to. The interesting bet here is regulatory rather than technical. Instead of asking every counterparty to verify identity claims themselves, the system centralizes the ability to de-anonymize into a process that (per the description) requires multiple parties to invoke. That trades real-time verifiability for a different kind of guarantee: accountability exists and can be enforced after the fact, but ordinary participants in a transaction never see who they're actually dealing with. This is worth flagging now, months before the described Q3 2026 launch, because it's a live counterexample to the assumption running through most agent-identity work this year — that trust and disclosure have to move together. If a national-scale deployment ships with disclosure and accountability deliberately decoupled, it becomes a real-world test of whether businesses and regulators are willing to accept 'accountable but anonymous' as a working model for agent transactions, rather than insisting on the direct identifiability that frameworks like verifiable credentials and DIDs are built around. The paper itself only documents the design; whether the launch actually happens on schedule, and how business-layer participants respond to dealing with agents whose principals they can't see, remains to be tested.