A healthcare multi-agent triage helper must never recommend specific prescription dosages, even if a model would. Built-in generic safety filters are not domain-specific enough. What should you implement?
Select an answer to reveal the explanation.
Short Explanation
Answer A. Beyond platform defaults, AI-500 covers custom guardrails for domain-specific constraints and testing them with synthetic data. Encode the no-dosage policy at the right intervention points and validate with synthetic clinical attacks. Blunt removal of all health knowledge over-blocks legitimate non-prescriptive help. User vigilance and beam search are not controls.
Full Explanation
Correct answer: A. Design custom guardrails for domain rules (here: no specific prescription dosages) and validate with synthetic tests. Place checks on outputs and, where needed, on tools that could fetch dosing tables.
B is incorrect: over-broad refusal destroys useful non-prescriptive guidance (hours, locations, general education) and is not precise policy engineering.
C is incorrect: users cannot be the primary safety control.
D is incorrect: decoding parameters do not enforce clinical policy.
Document guardrail ownership and update procedures when clinical policy changes.