Why twelve agents instead of one big model?

A single model does classification. The Burna AI engine does reasoning under constraint. Twelve agents lets each step enforce a narrow, regulator-defensible decision: term extraction does not also do attribution; attribution does not also do confidence calibration; confidence calibration does not also do regulatory encoding. Each agent constrains the next, and the output space narrows at every step. The cascading constraint architecture is the moat. A single-model wrapper hallucinates; the Burna AI engine structurally cannot produce a grade outside the defined CTCAE set, cannot skip citation, and cannot contradict its own upstream findings.