This quiz is interactive and needs JavaScript. Here are the questions for reference:
1. A user writes: "Just get me a human, I don't want to explain this again." The agent can see the issue is a single-click password reset it is fully authorized to perform. What is the correct action?
- A) Perform the reset first, since resolving it saves the human's time, then confirm
- B) Ask one clarifying question to be sure a human is really needed
- C) Escalate immediately with a handoff summary, without attempting the fix
- D) Escalate only if the user repeats the request a second time
2. Which of the following is one of the three valid escalation triggers?
- A) The customer sounds angry and frustrated
- B) The model's self-reported confidence is low
- C) The problem is highly complex
- D) The relevant policy is ambiguous or silent
3. Which signal is always the wrong basis for deciding to escalate a support case to a human?
- A) An explicit request for a human
- B) The agent's confidence score being low
- C) Repeated tool failures with no path forward
- D) A policy that does not cover the situation
4. A structured handoff summary must contain all five required fields. Which of these is NOT one of them?
- A) Customer ID and verified identity
- B) The full raw chat transcript
- C) Root cause
- D) Recommended action
5. A retry loop that appends specific validation errors back into the prompt is MOST effective for which failure, and typically resolves it in how many attempts?
- A) Missing information; 2-3 attempts
- B) Formatting errors (e.g. nested vs flat); 2-3 attempts
- C) Missing information; unlimited attempts until success
- D) Formatting errors; a single attempt only
6. For a hard policy limit like "never process a refund over $500," what is the architecturally correct enforcement?
- A) A more emphatic, all-caps system prompt instruction
- B) An application-layer intercept that blocks the tool call server-side
- C) Asking the model to double-check the amount before acting
- D) Fine-tuning the model to refuse large refunds
7. In a human-in-the-loop extraction pipeline, at what confidence level are extractions sent to automated downstream processing rather than the Human Review Queue?
- A) At or above 90%, using field-level confidence scores
- B) Above 50%, using one aggregate score
- C) At or above 90%, using one aggregate score
- D) Above 75%, using field-level scores
8. In The Escalation Handoff decision tree, what distinguishes the "Complex Policy Issue" branch from the "I want a human NOW" branch?
- A) The policy branch escalates faster
- B) The policy branch calls account-context tools (e.g.
get_customer) first before escalating - C) The explicit-request branch requires gathering account context first
- D) The policy branch never produces a structured summary
9. A customer sounding angry or frustrated is, by itself, a valid reason to escalate to a human.
10. Complexity alone is a valid escalation trigger, because a sufficiently hard problem should always be handed to a human.
11. When scope or policy is unclear, the minimal footprint principle says to prefer which kind of action, and what does the Constitution warn so that "do less" is not abused?
12. A billing agent hits Max Retries Exceeded trying to extract a complete author list, but the source only says "et al." and points to an external document it was never given. The customer has not asked for a human. What is the correct next step?
13. A customer is clearly upset about a duplicate charge. The agent acknowledges the frustration and offers to resolve it; the customer then says "no, I want to speak to a person." What is the correct behavior?
14. An agent escalates a duplicate-charge case and hands the human: customer_id: "CUST-847392", root_cause: "Duplicate charges due to gateway timeout.", amount: "$47.00 USD", recommended_action: "Approve refund for $47.00 USD and notify customer." Compared to forwarding the entire chat log, why is this the correct handoff?