Offline deployment moves a model from the research environment into an air-gapped production network. This page specifies the full transfer protocol, the sequence of operations, integrity verification, and acceptance gates required before a model serves traffic in isolation.
Deployment Flow
The pipeline below shows every stage from research artifact to production serving.
Transfer Sequence Diagram
The sequence below shows the interaction between systems and roles during the offline transfer.
Secure Transfer Package Contents
Each package contains:
- Signed model weights (see Secure Model Packaging)
- Release manifest with hashes and signatures
- Dependency bundle pinned to exact versions
- SBOM reference for audit
- Deployment configuration for DeploymentManager
- Acceptance test script and dataset sample
Integrity Verification
OfflineOperator verifies every package before import:
- Signature validation against the published public key
- SHA-256 hash comparison for every file in the manifest
- Dependency bundle hash check
- SBOM completeness check
Any mismatch aborts the import and triggers a security incident review.
Acceptance Test
Staging runs a fixed acceptance test suite before production approval:
- Smoke inference on reference prompts
- Latency check against Performance Validation thresholds
- Benchmark run on the Optimization Set
- Regression check against the current production baseline
The Hidden Test Set is never used during staging or acceptance. It is reserved for final audit and research evaluation only.
Monitoring in Air-Gap
Production models in air-gapped environments report metrics to an internal monitoring stack. Log egress is permitted only through a one-way transfer mechanism reviewed by security. No model weights, training data, or inference outputs leave the air-gap.
Rollback
If monitoring detects regression during production, DeploymentManager triggers the rollback protocol defined in Model Rollback. The previous version is restored from EnterpriseRegistry.