1. The electrical circuit example needs more explanation, and I think this will highlight some deep issues around representing things which are seemingly trivially graphs, as graphs in practice. In what sense is a bog-standard resistor directed? I assume the reason that the graph is directed is because current has a sign and in an undirected graph it becomes ambiguous which way the current is flowing (also you may want components like diodes). But the directed representation also has issues: "can current flow from 'Vdd'  to 'n0'?" should be immediately answerable from the properties of Vdd and its edges. There are other ways to represent an electrical circuit. One is as a directed graph but with incident edges recorded - but iiuc, this is excluded from the latest version of the paper. Alternatively, one could have a mathematical object, the name of which I actually don't know: it looks like an undirected graph, but where each partial edge has additional, unique, end-point data, as well as the common weight. Things like this are the reason why I think we need a broader group to look at this proposal (i.e. beyond SG19) and if we possibly can we should involve someone from the mathematics community. Otherwise there's a real danger we end up missing important insights.

2. My comment about the structure of the paper changing was a reference to previous comparisons with boost::graph. I'm sure these were in an earlier version, or am I misremembering? Either way, it would be very helpful to have a proper discussion of e.g. the move away from visitors.

3. Re. the definition of a graph, there needs to be a proper discussion about whether the paper's definition of graph is what some authors call a multigraph and whether it does/does not include loops. These things are mentioned, in passing, when introducing algorithms, but terminology needs to be properly established. 

4. I think we're trying to do too much in one go in this paper. I think a great first step would be to build on mdspan and try to standardize (or at least understand) what might reasonably be called an unstructured span. This could be represented as a vector of vectors or as a vector with some auxiliary storage indicating where the partitions fall. The point is that an unstructured span, with the right invariants, is an adjacency list. If we can understand unstructured span and its desirable api, I think this will be incredibly valuable guidance for what a standardized graph container might look like.

5. IIUC, this paper excludes pure connectivity graphs. These are incredibly helpful and, if I've understood correctly that they are not supported, would be a major omission. Another good reason, imo, to start with unstructured span!

6. I'm not convinced by the load api. We don't have a load api for vector etc. Moreover, would it not be preferable to have appropriate constructors?