PRODUCTION SCENARIO
A public-facing tutoring bot on a frontier model gets a steady stream of users trying to push it into producing disallowed content. The team wants a cheap first line of defense before each message reaches the main conversation.
Which approach meets that requirement?
Answering here is anonymous. Nothing is saved unless you sign in.
Show answer and explanation
Answer: Pre-screen input with a lightweight model such as Claude Haiku 4.5 and structured outputs
A harmlessness screen puts a lightweight model such as Claude Haiku 4.5 in front of the conversation and constrains its verdict with structured outputs, which keeps the check cheap and parseable. Screening after generation costs frontier tokens on every turn, and a hand-kept blocklist is reworded around in minutes.