The draft adds a named entry for the transformer family: encoder-decoder structure, self-attention, and the difference between encoder-only and decoder-only use.
What the base framework establishes
The public summary of the draft says the amendment adds an entry on transformer architecture and algorithm at 6.5.3.2.9. It describes encoder-decoder structures and self-attention, and explains that the architecture allows a high degree of parallel processing in training and deployment of large language models. It distinguishes encoder-only models, used for representation and embeddings, from decoder-only models, used for generative tasks such as summarisation and question answering. The final wording is not held and this page does not reproduce it. What the public record establishes is narrower and still useful: the framework now has a place to describe a transformer, and that place separates the two ways the family is used.
Why it matters to an AI management system
This is the distinction most AI system descriptions get wrong. A team says it uses a transformer, which says almost nothing about the task. Encoder-only use produces representations; decoder-only use produces text. The risk, the evaluation and the human oversight differ between them. The amendment gives the description a vocabulary for that difference.
The public draft adds material at 6.5.3.2.9. Its final wording is not held, so this page records what the draft summary states and does not present it as published text. New in the draft. The public summary describes the content; the final wording is not held and is not stated. The entry number comes from the published draft summary, not from the final text.
Evidence a reviewer samples
| Category | Artefact |
|---|---|
| Architecture note | Encoder-only, decoder-only or encoder-decoder, with the task it serves. |
| Attention and context record | The context window, the input format and any truncation rule. |
| Vendor model note | The provider, model version and whether the architecture is disclosed. |
One question to prepare for. Is the transformer used to represent or to generate, and how do you know?
Common gaps
- The description says only that the system uses a transformer.
- An encoder-only embedding model is described as generative because it sits near a language model.
- The context window and truncation rule are absent, so input limits are invisible to oversight.
Prepare before publication
Add an architecture field to the model card with the three options and the task. Record the context window and what happens to input beyond it.
Dates that matter
No separate date attaches to 6.5.3.2.9; the record dates are the stage date, the planning estimate and the CEN expectation stated above.
Cross-references
6.5 Tools · 6.5.3.2.10 Generative model · 7.5 Self-supervised machine learning
Source
Source for 6.5.3.2.9: ISO/IEC 23053:2022/FDAmd 1, new clause 6.5.3.2.9 (public draft summary). The public sample stops before 6.5.3.2.9, 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.