The room the model doesn’t see

Grounding a model in your material is only reassuring if you also know what that model is not being handed. The conversation surface is explicit about the boundary: a cloud model sees the slice of material relevant to the question in front of it, on that one call, and nothing else — not the rest of the base, not the graph in full, not a running transcript of everything asked before unless that history is itself part of what the question needs.

A locally-run model, by contrast, can be handed more without that boundary meaning anything different, because nothing leaves the machine at all. Which model answers a given question is a setting a person controls, and the surface is built so that switching that setting changes only who is doing the reasoning — never what the base or the graph are willing to keep, or who else gets to see it.

	A commercial model, called by name	A model running on this machine
What it sees per question	Only the relevant slice, sent for that one call	Whatever is handed to it — nothing ever leaves the machine either way
Who chooses it	The person, per question or as a default	The person, per question or as a default
What stays put regardless	The base and the graph, untouched by the choice	The base and the graph, untouched by the choice

That last row is the point of the whole surface: the model is a setting, not a foundation. Nothing about the durable material depends on which one happens to be answering today.