How does Burna AI prevent AI hallucinations?

Four structural defenses. First, the output space is hard-constrained to valid CTCAE terms and grades; a grade outside the set cannot be produced. Second, citation is mandatory; a grade without a source sentence and a CTCAE criterion cannot be produced. Third, each agent in the cascading constraint pipeline bounds the valid output space of every downstream agent; an answer that contradicts upstream findings cannot be produced. Fourth, every grade ships with confidence calibration so low-confidence outputs are routed to clinician review with priority. Human-in-the-loop is mandatory; the clinician reviews and approves before any grade becomes a record.