Question 240
An audit of a Cortex-enabled workflow finds the following issue: An internal drafting tool overblocks legitimate medical safety terms. What decision most accurately addresses the capability boundary in the scenario?
Explanation
Correct Answer
“Configure and evaluate supported guardrails around generated content, including clear escalation for flagged or blocked responses” is supported by this requirement: An internal drafting tool overblocks legitimate medical safety terms. Safety controls are defense-in-depth and must be measured for both under- and overblocking. Verify it using false-accept and false-block rates on tested harmful-output and benign-output samples, with escalation outcomes.
Incorrect Answers
“Hide safety failures from reviewers so the generated text appears reliable” would not meet the requirement that an internal drafting tool overblocks legitimate medical safety terms. Safety controls are defense-in-depth and must be measured for both under- and overblocking. The more reliable approach is to configure and evaluate supported guardrails around generated content, including clear escalation for flagged or blocked responses.
“Assume a friendly tone instruction prevents unsafe model outputs in every situation” would not meet the requirement that an internal drafting tool overblocks legitimate medical safety terms. Safety controls are defense-in-depth and must be measured for both under- and overblocking. The more reliable approach is to configure and evaluate supported guardrails around generated content, including clear escalation for flagged or blocked responses.
“Use only a final regular expression for profanity and declare all harmful content risks solved” would not meet the requirement that an internal drafting tool overblocks legitimate medical safety terms. Safety controls are defense-in-depth and must be measured for both under- and overblocking. The more reliable approach is to configure and evaluate supported guardrails around generated content, including clear escalation for flagged or blocked responses.