It adds a transformer architecture entry and a generative model definition, expands the model collapse note at 7.5, and updates the bibliography. It adds no management system requirement.
The public draft summary names four changes. The transformer entry describes encoder-decoder structures and self-attention, and separates encoder-only use for representations from decoder-only use for generative tasks. The generative model definition covers models that produce new data from a learned distribution and lists the families, including autoencoders, generative adversarial networks, normalising flows, transformers, diffusion models and autoregressive models. The 7.5 note warns that training on generated output can degrade quality across cycles. The bibliography adds an AI-generated content survey. The base structure, the pipeline and the annex remain.
Does the draft change the pipeline?
The public summary does not state a change to clause 8. The pipeline and its cross-cutting concerns remain the base framework.
Does it add a definition of large language model?
The public summary describes transformer use for large language models but does not state a separate LLM definition. The 22989 amendment is the terminology project to watch.
Where this comes from
The clause record states each unit in our own words and cites the source. The amendment page separates the public draft summary from the published base edition. The amendment record is ISO standard 88149.
Not legal advice; does not certify. The amendment wording is not held and is not stated.