Deployment packages the trained model for a runtime, which may be a service, a container or a device, and re-checks it when optimisation changes it.
What the base framework establishes
The clause covers putting a trained and evaluated model into an operating environment where it can serve production data. Models can be exposed through a web service or built into a native application. Packaging options include servers, clusters, virtual machines, exchange formats such as ONNX and NNEF, containers that carry their dependencies, and deployment on devices including drones. The runtime may differ from the development environment, and the clause warns that this difference needs care. Optimisation for a target platform, such as quantising weights, can change the model enough that a new evaluation is needed, sending the work back to verification.
Why it matters to an AI management system
Deployment is a change, and the framework treats it as one. A model that was evaluated in one runtime and optimised for another is not the same artefact. The record needs the deployed version, the runtime and the re-evaluation decision, or the evidence describes a model that never reached production.
The public draft summary names no change to 8.6 Model deployment. The final text could still touch it.
Evidence a reviewer samples
| Category | Artefact |
|---|---|
| Release record | The model version, the package format and the runtime. |
| Environment difference note | How production differs from the development environment. |
| Optimisation re-evaluation | The optimisation applied and the evaluation repeated after it. |
One question to prepare for. Is the model in production the exact model you evaluated?
Common gaps
- The deployed model cannot be traced to the evaluated version.
- Quantisation is applied for speed and the evaluation is not repeated.
- The runtime difference is discovered only when behaviour changes.
Prepare before publication
Tag the release with the evaluated model version and record the runtime. Add a re-evaluation gate for any optimisation that changes the weights or the arithmetic.
Dates that matter
No separate date attaches to 8.6; the record dates are the stage date, the planning estimate and the CEN expectation stated above.
Cross-references
8.5 Verification and validation · 8.7 Operation · 6.1 Overview of the machine learning system
Source
Source for 8.6: ISO/IEC 23053:2022, clause 8.6. The public sample stops before 8.6, so this page relies on the public contents and the public summary and does not quote the clause. The two official records are linked on this page so any line can be checked.
Not legal advice; does not certify. A reading of the framework and the evidence a reviewer commonly samples, not a substitute for the standard or your auditor's rules.