Air-Gapped AI Deployment Checklist

The AI application, model runtime, vector store, UI, logs, and tools can operate without outbound internet access.

External models, packages, datasets, documents, and updates enter through a controlled staging area before approval.

Artifacts are scanned, reviewed, approved, signed, and recorded before being moved into restricted environments.

Models, containers, Python wheels, prompts, tools, and configuration are stored in an internal registry or repository.

Every model, package, container image, prompt, and tool has a version, checksum, source, and approval record.

Applications can be built and redeployed using only approved internal artifacts.

Ollama, vLLM, llama.cpp, OpenShift AI, or another approved runtime is installed and validated.

Documents, embeddings, chunking, retrieval, citations, and vector storage stay inside the approved boundary.

User, prompt, model version, retrieved sources, tool calls, approvals, and outputs are traceable.

Users, admins, tools, document collections, and model operations follow least-privilege access.

Latency, token volume, model errors, retrieval failures, GPU utilization, and audit events are monitored.

A failed model, prompt, package, or tool update can be reverted using known-good approved versions.