A language model is asked for a description of the section in a fixed JSON format, never for markup. A validator decides what of its answer is allowed through, and one renderer draws every result: layered views in CSS, or a 3D object assembled from simple solids. Pictures are generated on request. The JSON is the deliverable, and it renders anywhere the motion kit does.
A model that writes markup can write anything, including a link that runs a script. A model that fills in a form can only fill in the form. The scene format is that form, so the page decides what is possible and the model only chooses.