Exams / AIP-C01 / Study guide / Transparency and explainability (citations, reasoning traces)
Transparency and explainability (citations, reasoning traces): AIP-C01 study notes and practice questions
What to know about Transparency and explainability (citations, reasoning traces) for the AIP-C01 exam, from the AI Safety, Security, and Governance domain (20% of the exam guide), then 1 free question to test it. Part of the AIP-C01 study guide.
AIP-C01 asks how to make a generative system explainable to the people who rely on it: end users who need to check an answer, and reviewers who must understand a decision. The right mechanism depends on who is reading and what they need to verify.
Citations for end users
RetrieveAndGenerate returns citations that point to the source documents and passages behind an answer. Show the document name, location and a link where the user can open the original. Citations only help if they support the claim they are attached to, so evaluate citation precision and coverage with Bedrock RAG evaluations.
Traces for reviewers
Agent steps are recorded through AgentCore Observability, which emits OpenTelemetry traces with a span for each model call and tool call. Reviewers can see which tools ran, with what inputs and outputs. Keep these traces internal, because they can contain personal data or system details. Give end users a short summary of steps, not raw traces. Bedrock Agents Classic has its own trace feature, but it is closed to new customers.
Confidence and uncertainty
Show uncertainty without implying false precision. A grounding or relevance score can drive a “low confidence, please verify” label. A not-found answer is a valid, honest result. Do not present a raw model score as a calibrated probability, because it is not one.
Disclosure and provenance
Tell users when content is AI-generated, in the interface and in exported documents. Amazon image models such as Nova Canvas add an invisible watermark to generated images, which supports provenance checks. Other providers’ models may not, so keep your own records too. Keep a record that links each published item to the model and prompt version used.
Evidence panels
An evidence panel shows the answer next to its supporting passages, with links to the sources. A reviewer can scan the claims and the evidence side by side, which speeds up approval in regulated workflows.
How to choose
Scenario wording
Choose
“Users must check where the answer came from”
Citations from RetrieveAndGenerate with links to sources
“A reviewer needs the steps an agent took”
AgentCore Observability traces, kept internal
“The answer may be uncertain, so say so”
Confidence label from grounding or retrieval scores
“Content must be marked as machine generated”
Interface disclosure, plus provenance records
“Approvers check claims against sources”
Evidence panel with passage-level links
Common traps
Showing citations that do not support the sentence next to them. Check citation coverage, not only whether a citation exists.
Exposing full agent traces to end users, which can reveal internal data.
Treating a model’s own explanation as proof. A model can explain a wrong answer convincingly.
Presenting a raw score as a percentage of truth.
Using Bedrock Agents Classic traces for a new build, since that service is closed to new customers.
A construction firm runs an assistant that answers site safety questions for its field crews. The assistant uses the firm's own safety manuals as its source.
Crews often ask whether an answer is official before they act on it. The firm wants each answer to show the manual and section it came from, so that a supervisor can check it on the spot.
The assistant must still return one written answer per question, not just a list of text passages.
Which approach meets these requirements?
Choose one answer.
AA. Prompt the model to end each answer with the phrase according to the safety manual, without giving the crew any link, document name or section reference.
BB. Call the RetrieveAndGenerate API, which returns a generated answer with citations to its source passages, and show those citations next to each answer.
CC. Call the Retrieve API only, and display the matching passages to the crew as raw text, leaving the crew to read the passages and decide the answer for themselves.
DD. Turn on model invocation logging so that each prompt and answer is saved, and let supervisors look up the saved records after an incident has occurred.
Show the answer and explanation
Answer: B
A. Wrong: A phrase like this claims a source without pointing to one, so a supervisor still cannot verify the answer. The model can also invent a source name, which defeats the purpose of showing where an answer came from.
C. Wrong: Retrieve returns relevant passages with their locations, but it does not generate an answer, so the assistant would not give the one written answer the firm requires. Crews would have to interpret the manuals on their own. Learn more: Amazon Bedrock Knowledge Bases (opens in a new tab)
RetrieveAndGenerate returns a generated answer with citations to the passages it used, so the interface can show the source next to each answer. Retrieve alone does not produce an answer, prompting for a phrase creates an unverifiable claim, and invocation logs only help after the fact.
Want more on Transparency and explainability (citations, reasoning traces)? The 5 full-length practice sets have 6 more Transparency and explainability (citations, reasoning traces) questions, 375 in all, none of them in the free exam. One payment of $14.99, no subscription. See the practice sets →
Study Transparency and explainability (citations, reasoning traces)