How does Burna AI prevent AI hallucinations in adverse event grading?

By constraining what the engine can say and making every claim checkable. Terms resolve against the CTCAE catalog for the trial's version and grades against that term's published criteria, so output stays within valid CTCAE terms and grades. Every grade carries the source sentence and the criterion it meets, and a grade whose citation set is incomplete is flagged before signature. Each of the twelve agents bounds what the next may conclude, so a step cannot contradict an earlier finding. Every grade carries a confidence, and the engine keeps what it first proposed, so an override stays visible. When the trial, enrollment or regimen for a patient cannot be resolved, the engine stops rather than guessing. No grade becomes a record until a qualified investigator reviews and signs it.