We publish the verification layer for AGI.

When Agents Need the Real World: Verifiable Human Execution for Reality-Bound Agent Tasks

A System Model, Prototype, and Evaluation Plan for AI2Human

Autonomous agents can browse, call APIs, write code, and coordinate digital workflows — but they still fail when a workflow requires someone to enter a store, inspect a shelf, verify a delivery, collect a signature, or produce fresh evidence from a physical location. We introduce Structured Human Execution Infrastructure: a system layer that lets agents turn a blocked real-world step into a bounded human-executable task, receive structured proof, verify the result, and release payment only after acceptance. The contribution is an agent-native interface in which tasks, proof requirements, verification state, settlement state, and audit logs are machine-readable and can be reintegrated into an agent workflow.

Agents can't prove they did the job — so we built the verify primitive

Agents are great at claiming things. We built verify(claim, evidence) → verdict + receipt — the missing primitive that turns “done” from a model’s opinion into a verifiable fact.