PRODUCTION SCENARIO
A public-facing tutoring bot running on a frontier model gets a steady stream of users trying to push it into disallowed content. The team wants a cheap check 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 each message with Claude Haiku 4.5 and constrain its verdict with a schema
A harmlessness screen puts a lightweight model such as Claude Haiku 4.5 in front of the conversation and uses structured outputs to constrain its verdict to a parseable classification, which is what 'cheap, and before the message lands' means. A hand-kept blocklist is reworded around. Throttling repeat offenders is a real mitigation on the same page, but it responds after the fact rather than screening the message. And checking the finished answer spends frontier tokens on every turn.