Knowledge graphs should be projections, not operational truth

Knowledge graphs are compelling because they make relationships visible: a decision can be connected to its inputs, policies, artifacts, owners, and downstream effects. The danger starts when the graph is asked to become the source of truth for every operational interaction.

An operational system needs clear writes, well-defined consistency guarantees, and failure behaviour that its owners understand. Graph-shaped queries do not remove those needs. Folding every command path into a shared graph often introduces ambiguous ownership and an expensive coordination point.

Treat the graph as a projection instead. Let the authoritative systems publish explicit events or state changes. Build the graph from those contracts, with clear freshness and repair semantics. It can then be optimised for discovery, lineage, investigation, and impact analysis without becoming a hidden dependency for the core transaction.

This distinction also makes the graph more honest. A projection can show what is known, when it was observed, and where it originated. That is often more useful to an architect than a graph that looks complete while quietly carrying operational responsibilities it cannot safely own.