Microsoft AI-103 Practice Test 41
Question 400
An agent's cost rises and its tool decisions become unsafe despite a stable final-answer score. Which TWO monitoring additions address the hidden regressions? Choose TWO.
- Record only model output text while discarding the identity, cost and authorization outcome of the underlying tool attempts.
- Correlate retrieved document IDs and ranking scores with agent output quality
- Break down prompt and completion tokens by model invocation and agent step
- Assume model tool selection is safe whenever the final-answer quality score stays constant.
- Measure unauthorized tool-attempt and policy-denial events alongside task success
Explanation
Correct Answers
C:Step-level token accounting identifies costly loops and unnecessary context. Token-by-step profiling detects cost growth and policy-attempt telemetry detects unsafe intermediate behavior.
E:Outcome success can hide risky intermediate attempts that require security review.
Incorrect Answers
A:Text-only telemetry cannot expose which agent step drove cost or whether a forbidden tool action was attempted. Record token costs and policy-denial events independently.
B:Coupled retrieval and answer telemetry isolates the evidence-selection stage. That mechanism concerns connect search relevance problems to agent responses, whereas this case tests how to monitor step cost and unsafe tool execution.
D:Final-response ratings can conceal unsafe intermediate tool decisions; they do not substitute for tool-policy telemetry.
Question 401
A summarizer produces varied creative prose when customer reports require near-identical structured wording. What generation change helps?
- Separate trusted instructions from retrieved untrusted evidence and forbid instruction promotion
- Use lower sampling variability with a strict response template
- State an explicit output-length constraint and validate it after generation
- Split the report into additional creative subrequests with different free-form output styles but do not validate a common response schema.
Explanation
Correct Answer
B:Constrained generation and a defined schema reduce style variation for deterministic business tasks. Here the relevant capability must let the team reduce creative variation in structured summaries. An alternative targeting isolate prompt-injected external content solves another problem.
Incorrect Answers
A:Prompt structure should not let a third-party document override trusted application rules. This helps when a team needs to isolate prompt-injected external content; it does not establish how to reduce creative variation in structured summaries.
C:A measured output contract is more reliable than vague requests to be concise. That capability is appropriate for enforce compact output length, not this separate task of finding how to reduce creative variation in structured summaries.
D:More independently generated prose tends to increase variation. A consistent report requires constrained sampling, explicit template or structured schema, and validation of required fields and style.
Question 402
A public chatbot sometimes generates explicit harmful instructions when baited by malicious prompts. Which platform protection should be configured?
- Apply tested content-filter policies to unsafe categories of generated text
- Attach a provenance watermark to the response while leaving harmful instruction generation unfiltered.
- Evaluate false-positive and false-negative rates on representative safe and unsafe prompts
- Validate tool arguments and authorize every side effect on the server
Explanation
Correct Answer
A:Safety filters can block or restrict prohibited outputs independent of model fluency. Here the relevant capability must let the team moderate unsafe generated content. An alternative targeting label synthetic media provenance solves another problem.
Incorrect Answers
B:Origin labels cannot block harmful content. Apply and evaluate suitable content safety filters at the response boundary.
C:Safety tuning must balance missed harms against overblocking legitimate users. That capability is appropriate for measure moderation precision and recall, not this separate task of finding how to moderate unsafe generated content.
D:A prompt-only instruction is not an access-control mechanism; programmatic checks constrain sensitive writes. Here the needed outcome is to moderate unsafe generated content; an option designed to enforce input and authorization policies at tool calls would not directly provide it.
Question 403
A model returns a caption for a clip but fails to cite which segment supports its claim. What metadata is required?
- Publish a free-form clip description without retaining source frame timestamps or segment identifiers.
- Request object-region localization with coordinates for the source frame
- Attach segment time ranges to observed claims and extracted visual evidence
- Compare temporally sampled frames while preserving their order
Explanation
Correct Answer
C:Timestamped evidence allows reviewers to verify what actually happened in video. In this scenario the team must ground video summaries in timestamped segments. A feature used to balance frame processing cost against missed events does not substitute for it.
Incorrect Answers
A:A caption can be plausible but untraceable. Link each observed visual claim to the source clip and the precise time range that supports it.
B:Region or bounding-box outputs identify the location, not just presence, of an object. Its primary purpose is to locate a detected subject in a frame; the stated constraint instead demands a way to ground video summaries in timestamped segments.
D:Visual reasoning about motion requires correct temporal sequence, not unordered snapshots. This helps when a team needs to compare observed visual changes through time; it does not establish how to ground video summaries in timestamped segments.
Question 404
A TTS application pronounces a medication name incorrectly for patient instructions. What should the team adjust?
- Speak every unknown medication name using the default grapheme pronunciation.
- Increase the audio sample rate without inspecting the name pronunciation.
- Supply supported pronunciation lexicons or SSML pronunciation controls and verify output audio.
- Edit only the response's font and punctuation and assume spoken output improves.
Explanation
Correct Answer
C:Custom pronunciation controls can disambiguate unfamiliar names, but generated audio still needs safety-sensitive review.
Incorrect Answers
A:Default pronunciation may repeat the hazardous name error in patient instructions.
B:Higher sample fidelity does not correct phoneme selection for a specialized term.
D:Text formatting does not provide the correct phonetic representation for TTS.
Question 405
A document parser extracts a bank statement amount but the minus sign disappears during OCR. What validation protects the financial value?
- Accept the extracted amount because all digits except the sign were recognized.
- Compare extracted amounts to original page regions and validate signs, totals and currency.
- Round every extracted value to the nearest positive whole number.
- Recheck only the document title rather than the original amount region.
Explanation
Correct Answer
B:OCR can introduce digit and sign errors; arithmetic and visual checks detect critical extraction failures.
Incorrect Answers
A:A missing minus sign reverses the financial meaning even when the digits match.
C:Rounding and forcing positivity can conceal negative transactions and destroy reconciliation.
D:Title metadata does not validate the sign, currency or arithmetic relationships of the field.
Question 406
A model-call trace contains total tokens but the team cannot tell which tool chain is expensive. What attribution should be added?
- Aggregate all requests into one monthly token total without agent-step labels.
- Measure only the customer's browser rendering latency.
- Aggregate token and latency cost by model, agent step, tool span and request correlation ID.
- Tag costs only with Azure subscription without model and tool identifiers.
Explanation
Correct Answer
C:Per-step attribution identifies expensive orchestration paths and supports targeted optimization.
Incorrect Answers
A:A global total cannot identify the expensive tool chain or specific execution step.
B:Browser timing does not attribute provider tokens or orchestration tool costs.
D:Subscription-level totals are too coarse to isolate expensive agent branches.
Question 407
A document-retrieval agent follows the words 'ignore previous rules' found inside a PDF and sends customer details externally. What threat needs a boundary?
- Promote instruction text extracted from an untrusted PDF to the agent's highest-priority tool policy.
- Detect and redact protected personal data according to policy before response delivery
- Validate tool arguments and authorize every side effect on the server
- Treat retrieved document instructions as untrusted data and enforce tool permissions
Explanation
Correct Answer
D:Indirect prompt injection attempts to turn lower-trust evidence into commands; tool authorization must not depend on retrieved prose. This directly supports the need to isolate lower-trust instructions from privileged actions, unlike a mechanism primarily intended to moderate unsafe generated content.
Incorrect Answers
A:Retrieved content is evidence, not authority to change system constraints. Treat embedded commands as untrusted data and enforce authenticated tool permissions independently of the document.
B:Sensitive-data handling requires explicit detection and release rules rather than relying on generic politeness. That capability is appropriate for prevent disclosure of sensitive identifiers, not this separate task of finding how to isolate lower-trust instructions from privileged actions.
C:A prompt-only instruction is not an access-control mechanism; programmatic checks constrain sensitive writes. Here the needed outcome is to isolate lower-trust instructions from privileged actions; an option designed to enforce input and authorization policies at tool calls would not directly provide it.
Question 408
An analysis pipeline stores frame embeddings without links to source footage. What indexing change is needed?
- Store every frame embedding under a newly generated identifier with no mapping back to the source recording or time offset.
- Associate extracted frame features with source video IDs and segment offsets
- Attach segment time ranges to observed claims and extracted visual evidence
- Choose temporal sampling based on event duration and risk of missing brief actions
Explanation
Correct Answer
B:Provenance links allow retrieved frame evidence to be located in the original recording. This resolves the specific need to map search matches back to source video, while a measure that helps find an event in a large video by timestamp addresses a different requirement.
Incorrect Answers
A:An isolated embedding cannot direct a reviewer to the corresponding evidence. Preserve stable video IDs and segment or frame offsets with indexed features.
C:Timestamped evidence allows reviewers to verify what actually happened in video. This helps when a team needs to ground video summaries in timestamped segments; it does not establish how to map search matches back to source video.
D:Sparse sampling can miss transient events, so interval selection must reflect task demands. The function serves to balance frame processing cost against missed events, whereas the actual requirement is to map search matches back to source video.
Question 409
A multilingual agent must refer to the exact turn where an issue was reported. What transcript feature is useful?
- Keep speaker-labeled timestamps aligned with recognized utterances
- Discard utterance timestamps once text has been translated
- Provide time-coded transcript excerpts and uncertainty flags
- Use an audio-capable multimodal reasoning workflow with explicit task criteria
Explanation
Correct Answer
A:Time and speaker grounding makes later claims traceable to specific audio segments. The decisive difference is whether the design can locate statements by speaker and time; the competing function serves to avoid overconfident emotion attribution from audio.
Incorrect Answers
B:Without time-aligned turns the agent cannot locate when an issue was mentioned. Speaker-labeled timestamps preserve provenance across language transformations.
C:Grounded transcription references let reviewers revisit contested acoustic evidence. Its primary purpose is to trace audio reasoning back to evidence; the stated constraint instead demands a way to locate statements by speaker and time.
D:Voice intent may depend on acoustic as well as lexical evidence; the model should be tested accordingly. That is relevant to interpret spoken intent using audio context, but the present scenario requires an ability to locate statements by speaker and time.
Question 410
A document contains signatures and checkboxes relevant to review. What content model should be considered?
- Use document analysis capable of representing detected layout elements and regions
- Treat a document's signature image and marked checkboxes as irrelevant decoration
- Use layout-aware extraction with table rows, columns and numeric fields
- Recover reading order based on detected document layout
Explanation
Correct Answer
A:Non-paragraph marks can be business evidence and need structured representation where supported. Here the relevant capability must let the team capture non-textual form elements. An alternative targeting locate extracted document fields in the source solves another problem.
Incorrect Answers
B:Signatures and checked fields can be material evidence in a review. An appropriate layout-capable representation should preserve these elements.
C:Flat OCR text cannot reliably preserve which prices belong to which line items. That capability is appropriate for extract invoice table layout and numeric values, not this separate task of finding how to capture non-textual form elements.
D:Layout-aware reading order is necessary for coherent extracted sentences across columns. Here the needed outcome is to capture non-textual form elements; an option designed to preserve multi-column reading sequence would not directly provide it.