Week 20: Quiz (10 questions, 8/10 to pass)

1. MCQ: Which one-line rule captures the Bedrock vs SageMaker split? (see Concepts §Bedrock vs SageMaker)

   • A) Bedrock = consume models; SageMaker = own models
   • B) Bedrock = train; SageMaker = call APIs
   • C) They are the same service
   • D) Bedrock requires GPU clusters

2. MCQ: The Converse API is described as the cleanest of the three clouds because… (see Concepts §Converse API)

   • A) It is the only API that supports images
   • B) One provider-neutral message/tool-calling format lets you swap the modelId and nothing else
   • C) It is free
   • D) It auto-fine-tunes models

3. MCQ: An AccessDeniedException on Bedrock usually means… (see Concepts §How it breaks)

   • A) Your Python is out of date
   • B) You haven't enabled the model in Model access, or your IAM policy lacks the action
   • C) The region is wrong
   • D) The model ID is too long

4. Short answer: In notebook 01, which two metrics does the RAG section compute, and what does each mean? (see notebook 01 cell 9 to 11)

5. MCQ: Bedrock model IDs like anthropic.claude-3-5-sonnet-20241022-v2:0 carry dated suffixes because… (see Concepts §Model catalog)

   • A) They are random
   • B) Models get superseded often, so you copy the current ID from the console rather than memorize it
   • C) They encode the price
   • D) They encode the region

6. MCQ: Which Guardrail filter masks names, addresses, and tracking identifiers? (see Concepts §Guardrails table)

   • A) Denied topics
   • B) Content filters
   • C) PII redaction
   • D) Custom word filters

7. Short answer: In notebook 02's dry-run local_guardrail, what two conditions cause a prompt to return GUARDRAIL_INTERVENED? (see notebook 02 cell 6)

8. MCQ: Provisioned Throughput is a bad idea for a lab week because… (see Concepts §Pricing)

   • A) It is pay-per-token only
   • B) It reserves model units per hour and bills whether you use them or not
   • C) It requires a credit card
   • D) It only works with Nova

9. Short answer: What does the Week 20 portfolio deliverable (the three-cloud matrix) require in every cell, and why does that make it a procurement decision rather than a blog post? (see Concepts §The problem / How it breaks)

10. MCQ: In notebook 02's dry-run, the final cell prints… (see notebook 02 cell 11)

    • A) RECALL: 1.000 GROUNDEDNESS: 1.000
    • B) BLOCKED: 3 ALLOWED: 2
    • C) EVAL_SCORE: 1.000
    • D) BLOCKED: 2 ALLOWED: 3

Answer key

1. A: Bedrock is the managed foundation-model layer (consume); SageMaker AI is the build/train/deploy-your-own layer (own).
2. B: The Converse API's single format across Claude/Nova/Llama/Mistral means you swap modelId and nothing else.
3. B: AccessDeniedException means the model isn't enabled in Model access or the IAM policy lacks the action, check both before touching code.
4. Recall: the fraction of questions where retrieval surfaced the correct source document; groundedness, the fraction of answers containing the ground-truth fact (cited, not hallucinated). Both are fractions over the question set.
5. B: Dated suffixes get superseded often; copy the current ID from the console, never memorize it.
6. C: PII redaction masks names, addresses, and tracking identifiers.
7. Either the prompt contains an insult word (idiot, garbage, stupid, useless) or it contains no freight keyword (shipment, refund, track, freight, delivery, policy, pallet, bill).
8. B: Provisioned Throughput reserves model units per hour and meters while idle; on-demand is the right choice for a lab.
9. Every cell must trace to a measured number from Weeks 18 to 20 (same golden set, seed, and metric). With per-cell evidence the matrix supports a recommendation; without it, it is just opinion, hence a procurement decision, not a blog post.
10. B: BLOCKED: 3 ALLOWED: 2 (three blocked prompts, insult, cake recipe, pirate joke, and two allowed freight prompts).