Exam-style
A team plans to give one Bedrock guardrail five jobs: block violent content, block questions about competitor pricing, mask phone numbers in responses, flag answers the retrieved passages do not support, and produce the quarterly bias report. Which job will the guardrail not do?
Reveal the answer
E. Producing the bias report, which is an offline measurement job rather than runtime enforcement
Guardrails covers the first four directly. Content filters score every prompt and response across the six harmful-content categories: Hate, Insults, Sexual, Violence, Misconduct and Prompt Attack. Denied topics block the competitor-pricing subject. Sensitive information filters detect the PHONE entity and either block the content or mask it with ANONYMIZE. The contextual grounding check flags responses the retrieved source does not support. Measurement sits outside all of that, so the bias report comes from an offline run: the Builtin.Stereotyping metric in an Amazon Bedrock evaluation job, or the prompt-stereotyping evaluation in the open-source fmeval library.
Q. What does Bedrock Guardrails actually enforce at runtime?
A. Six policy types, all applied at inference. Content filters score Hate, Insults, Sexual, Violence, Misconduct and Prompt Attack. Denied topicsSubjects you describe in plain language that a Bedrock Guardrail refuses to discuss, whichever way a user phrases the request. and word filters block subjects and exact phrases you name. Sensitive information filters block or mask PII. Contextual grounding checks flag responses the source passages do not support. Automated Reasoning checks return findings in detect mode rather than blocking. Guardrails enforces safety; it does not measure bias.
Why? Guardrails applies policy at inference time. Bias measurement is a separate offline job.