Guardrails and Output Moderation Quiz

Answers and explanations appear as you take the quiz.

1. What are guardrails in the context of a language model?
   • Safety checks that keep inputs and outputs within acceptable bounds
   • Decorative borders around the chat window
   • A way to make the model respond faster
   • Hardware that cools the servers
2. What does 'output moderation' specifically inspect?
   • The electricity used by the model
   • The text the model produced, before it reaches the user
   • The user's typing speed
   • The color scheme of the interface
3. Which is an example of an input-side guardrail?
   • Translating the answer to Spanish
   • Making the font larger
   • Filtering a prompt-injection attempt before it reaches the model
   • Adding citations to the answer
4. Why is moderating streaming (token-by-token) responses harder?
   • Streaming uses more colors
   • Streaming disables all classifiers
   • Users cannot read streamed text
   • Text is shown as it is generated, so harmful content can appear before a full check
5. What is the key balance guardrail engineers must strike?
   • Catching harmful output while minimizing false positives on legitimate requests
   • Making the model as slow as possible
   • Blocking every single response
   • Removing all logging
6. What is a privacy-conscious way to log moderation decisions?
   • Store every full conversation indefinitely with public access
   • Record only necessary information, with defined access and retention controls
   • Disable all privacy controls to simplify debugging
   • Publish blocked prompts so users can inspect them
7. What should pre-deployment guardrail testing include?
   • Only the prompts that already pass
   • No testing of borderline requests
   • Representative harmful and benign cases to measure misses and false positives
   • A single test prompt for every policy category
8. Why monitor guardrails after deployment?
   • Monitoring makes the model intrinsically safe
   • A deployed classifier never changes its error rate
   • It eliminates the need for user reports or evaluation
   • Usage and threat patterns can change, revealing new errors or policy gaps