What's Wrong with Semantic Bridging?

Semantic layers have grown from shared business vocabulary into architectures that bind concepts to metrics, relationships, grain, aggregation behavior, materialization, lineage, and increasingly AI context. This essay calls that architectural tendency "semantic bridging" and argues that it leaves a distinct jurisdiction of analytical identity and law scattered across the data stack. Using a 48-store missing-evidence case and an inventory-over-time case, it separates business meaning, governed analytical data, and material computation. The essay gives modern semantic systems full credit for real safeguards such as fan-out protection, symmetric aggregation, additivity controls, and retained aggregate state, while asking what analytical objects those safeguards protect and where their authority comes from. AI agents make the boundary urgent by removing the mandatory human stops that historically carried many of these distinctions. The proposed architectural test is simple: determination must remain distinct from execution, and a computable result should become an answer only when the analytical object that was asked for has actually been established.