How is this different from a single LLM with retrieval?

A single LLM with retrieval is one model and one prompt with a citation step bolted on. The Burna engine is twelve specialized agents in a cascading constraint pipeline. Each agent's output bounds the valid output space of every downstream agent, terminating at a citation-bound CTCAE grade that a clinician signs. The engine structurally cannot produce a grade outside the defined CTCAE set, cannot skip citation, and cannot contradict its own upstream findings. Constraint at the architecture level, not the prompt level.