Safety & limits
AI guardrails (Research Assistant)
Three layers of safety so the AI can never recommend, rank, or dose.
The Research Assistant is powered by Calanthe (NxGen AI Labs' flagship AI). It narrates a rule-based assessment — it does not interpret a patient's record from scratch. Three independent safety layers sit in front of it:
- System prompt. The model is given a research-only role with explicit prohibitions — no patient-directed language, no recommendations, no rankings, no dosing, cautious phrasing only.
- Output filter. Every generation is regex-checked. If banned phrasings appear, the generation is rejected.
- Deterministic fallback. When the host is offline OR the output was rejected, a template-built restatement of the assessment is returned. The user always gets something safe.
Calanthe runs as a dedicated agent for Teloma — its prompt, knowledge base, and safety guardrails are tuned for cancer research narration and shared with no other product.