It has been reported by @anoe that Order2 (O2) Graphs sometimes have a problem: they do not show edges between nodes. Alexandre thinks this is due to a bug in the distributional similarities matrix algorithm, where some numbers ends up being 0 or negative due to the fact a value is too small to yield a positive log.

Alexandre will try to revive the old reference implementation we had for the O1 and O2. The plan is to fix the immediate issue but also work on testing the feature, either by writing tests which checks we will always end up with edges (i.e. we won't generate matrixes with all zeros) but also that the values produced by the reference and actual implementation agrees, so that we can express our high-level invariants in the easy reference implementation, and later reflect that into the fast (and somewhat low-level) Accelerate one.