I think you're describing something like a hybrid social network/semantic web with versioning, knowledge provenance, and asynchronous peer review.
One serious problem for any such system is ontology selection: how is one to represent the entire body of scientific knowledge under a single type system? Different fields of inquiry make use of extremely diverse conceptual models. I suppose mathematics are in a way a unifying language, but there's hardly a single homogeneous mathematical discipline.
The present "weakly" connected network has almost zero technical barriers to entry. It uses well-established technology within a well-established workflow, and it offloads the hard, fuzzy work (e.g., all the model-binding that would presumably take place in the proposed system) to the most flexible computing device we know of: the brain. Everything is already freely downloadable/usable, for the most part (lots of research is open access, and what isn't can often be obtained from the investigators by request).
That said, maybe the sort of thing you describe could be translated into a research question. One could try to compare the shape of various data under different encodings, for instance (some sort of topological analysis?) to identify similar structure? I think category theory has been used to unify previously disparate regions of mathematics.
There are already a few entries in the social network/resource-sharing platform space. Have a look at Open Science Foundation. Academia and ResearchGate are similar, but without the materials-and-data-sharing.