Page 3

Every field note remains reachable as the lab grows.

A failure-injection drill can test identity, reconnection, retained state, idle shutdown, and clean restart across an agent runtime boundary.

Test shared chat clients against event-loop, thread, session, and streaming boundaries before approving a Python agent-host topology.

OpenAI’s research-intern milestone separates agent runtime, spending, activity, and research progress into distinct measurement layers.

Review behavioral control and operational observability separately when approving an agent deployment.

Document the physical control boundary, evidence, and safe recovery conditions before an AI agent receives write access.

Compare AWS and Google Cloud’s documented agent-governance scope without mistaking vendor descriptions for operating evidence.

A mixed-version drill separates protocol negotiation, effective capabilities, and session recovery before an agent-host rollout.

Fable 5.1 and Mythos 5.1 show why model procurement now has to capture access, safeguards, routing, retention, and cost.

A review gate separates benchmark gains in automated alignment research from evidence that a method will transfer to another model or setting.

A proposed escalation ledger links each cloud approval to its workflow step, disclosed payload, destination, and returned result.