This is the Implementation Edition (Book Two, "Architecture and Verification") for architects and engineers, distinct from the Foundations Edition. Here's a back cover description in your voice, sharpened against that volume's actual scope.
THE THREE-BODY PROBLEM OF CLASSIFIED AI Implementation Edition - Architecture and Verification
An Authorizing Official does not sign architecture. The AO signs evidence.
The Foundations Edition defined the eight properties a classified AI system must satisfy to be authorized. It did not tell you how to make them true, or how to prove they stayed true after the assessor went home.
This volume does.
Architecture and Verification is the engineering half of the framework. It builds the operational trust layer one component at a time: the Delegation Token that binds every agent action to a human and an intent, the Unified Activity Record that leaves nothing unlogged, the Agentic RAG gateway that enforces classification at the point of retrieval, and the Governance Engine that gates tools by deterministic policy rather than model instruction. Then it builds the accountability layer the AO actually reads: the evidence stream, policy-as-code, the risk register, and the operating picture.
Part Three goes past the baseline. Policy-gated autonomy. Automated ATO pre-submission scanning. Cross-system federated trust. The mosaic effect. Runtime detection and posture. The verification stack that turns "trust but verify" from a slogan into a control.
One rule runs through all of it: if the control ran, evidence exists; if evidence does not exist, the control never ran.
A less capable agent with a well-designed harness outperforms a more capable agent without one. The harness is the product.
Written for system architects, developers, security engineers, and DevSecOps teams who build what the AO will authorize, and for the AO who has to read what they built. Read the Foundations Edition to understand the problem. Read this one when you are ready to build.
