Question 340
At Summit Pharma, a regulated information center is improving product information during an agent routing evaluation (SF-A10-01). The administrator observes that an agent sometimes sends cancellation requests for orders that are already shipped. Any fix must ensure that sensitive product claims must cite approved knowledge. Which observable result would best demonstrate that the correction actually worked?
Explanation
Correct Answer
This is the direct acceptance signal for the reported failure: In a controlled Product2 acceptance test, a shipped-order test cannot invoke cancellation while eligible orders can; the record is retained under Summit Pharma's rollout (SF-A10-01). The chosen correction was to filter the cancellation action on the required order state before execution; the measurement therefore tests the relevant source, routing or authorization outcome rather than unrelated activity.
Incorrect Answers
The proposed observation “An attempted cross-account request fails despite a plausible supplied email address; confirm this and retain a reproducible test record” would check a different original case: An agent confuses a user-provided email address with the authenticated contact ID. The correction must ensure that financial identifiers are masked outside authorized roles. Its mechanism is: Trusted linked identity is not interchangeable with unverified conversational text. In the present case, an agent sometimes sends cancellation requests for orders that are already shipped. The correction must ensure that sensitive product claims must cite approved knowledge; the more direct acceptance evidence is “A shipped-order test cannot invoke cancellation while eligible orders can; confirm this with production-equivalent source conditions” because action filters constrain which actions are available under defined conditions.
The proposed observation “Two consecutive credit requests produce at most one approved credit action; confirm this and compare with approved baseline behavior” would check a different original case: An agent issues a customer credit twice during one support session after a repeated request. The correction must ensure that safety-critical changes need an attributable human approval. Its mechanism is: A mutable state flag can enforce a rule independent of LLM wording. In the present case, an agent sometimes sends cancellation requests for orders that are already shipped. The correction must ensure that sensitive product claims must cite approved knowledge; the more direct acceptance evidence is “A shipped-order test cannot invoke cancellation while eligible orders can; confirm this with production-equivalent source conditions” because action filters constrain which actions are available under defined conditions.
The proposed observation “A user without field access cannot obtain the protected value through the agent; confirm this using a least-privilege test identity” would check a different original case: A template returns a confidential account field in a customer-facing response. The correction must ensure that policy terms come from the current approved revision. Its mechanism is: Field-level security still governs agent grounding and action results. In the present case, an agent sometimes sends cancellation requests for orders that are already shipped. The correction must ensure that sensitive product claims must cite approved knowledge; the more direct acceptance evidence is “A shipped-order test cannot invoke cancellation while eligible orders can; confirm this with production-equivalent source conditions” because action filters constrain which actions are available under defined conditions.