How is Burna AI different from asking a single LLM to grade adverse events?

A single large language model classifies under uncertainty; Burna AI reasons under constraint. Its engine is twelve specialized agents in a cascading constraint pipeline, from extraction and verification through causal attribution and confidence calibration, where each agent's output bounds what every later agent can return. Grades are restricted to valid CTCAE grades for the version the protocol names, and every grade must cite the line in the medical record it came from and the criterion it meets; a grade with an incomplete citation set is flagged before signing. Attribution uses the WHO-UMC and Kramer algorithms with the reasoning shown, rather than a free-text opinion. Two patents filed on the architecture. An investigator reviews and signs every result.