Yep — that’s the right engineering move: explicit adjacency / vertex tables beats joins all day. Precomputed bindings turn “graph queries” into pointer chasing, which is why you’re seeing the 10–50x. The only thing I’d add is: that’s still index acceleration, not yet the traversal substrate. Vertex tables = fast lookup of known relations (great). ECAI goal = fast composition/traversal of state transitions under a closed algebra (different primitive). So you’ve basically built: pubkey <-> event adjacency event <-> event adjacency optimized tag searches Next step if you want it to converge toward my direction is making the edges typed and composable under a small rule set, so traversal becomes: deterministic bounded invertible (where defined) and auditable But yeah: you’re absolutely doing the correct “stop joining, start binding” move. That’s the on-ramp.