I think of relays as independent servers, which do not gossip to other relays.

In the following paragraph, strings of capital letters are just monikers for imaginary relays.

I don't like structuring relays into some kind of network because then I have to trust more nodes than I strictly need to.  If I have a digitally signed message (kind 10002) from Steve Bellovin saying that he posts to relay XYZ, then why not go directly to XYZ to get his stuff?  Why should I trust that ABC will get it from DEF who will get it from GHI... too many points of failure.

I know that some people do trust LMNOP and will get everything from LMNOP just to offload workload and maintain privacy, and this model I call a "client proxy model" and it makes sense -- you just sort-of extend your client forward where the forward part goes to XYZ and your local part just talks to the forward part.  But I prefer not to proxy my client's work, I prefer direct access, getting notes from the source, with the minimum chance of censorship because I'm on a desktop and I can.