Question 400 A model's output safety changes after a deployment update. What release record enables root-cause analysis? Only the current latest label Only the team member who noticed the issue Immutable model, prompt and safety-policy version references in traces Only current moderation thresholds without recording their version at request time Explanation Correct Answer C:Version provenance permits correlating regressions with exact effective settings. Incorrect Answers A:A mutable alias hides which version served earlier requests. B:Reporter identity is not the effective service configuration. D:Current policy values cannot establish which filters actually applied to earlier decisions.
Question 401 A finance assistant proposes a refund after reading an untrusted PDF. Which TWO checks must precede payment execution? Require authorization linked to the exact payee, amount and operation Execute any command appearing as bold text in the PDF Retry failed refunds using new keys without reconciliation Trust the model's confidence score as approval Validate refund eligibility and amount against trusted transactional records Explanation Correct Answers A:Approval must be bound to the actual irreversible business action. E:A document's text does not prove a refund is permitted. Incorrect Answers B:Source documents are lower-trust content and may contain prompt injection. C:Blind replay can create duplicate financial effects. D:Confidence is not an external authorization mechanism.
Question 402 The team needs to identify which production requests triggered policy-based human approval. What should traces preserve? Keep the approval ticket number without a reference to tool input or actor Record the model recommendation but omit the external approval result The action intent, policy rule, approval ID, actor and resulting decision Store only the final successful tool response without policy context Explanation Correct Answer C:Decision provenance supports accountability without exposing irrelevant secrets. Incorrect Answers A:A bare ticket cannot prove which actual action was authorized. B:Model advice does not demonstrate that a required policy gate ran. D:Success does not demonstrate that this operation had been approved.
Question 403 A search index was updated but response relevance dropped sharply only for French content. What should be inspected? Only the number of English portal users Only pricing-tier label capitalization Language analysis, embeddings, preprocessing and evaluation samples for French records Only deployment plan cost tags Explanation Correct Answer C:Language-specific ingestion mistakes can reduce recall without affecting English results. Incorrect Answers A:English usage does not explain French retrieval failures. B:A label does not validate tokenization and embeddings. D:Cost allocation metadata is separate from search relevance.
Question 404 A Foundry project endpoint starts returning 429 errors after new traffic. What deployment-level remedy should the team assess? Capacity planning, quota limits and controlled request scheduling Remove all safety evaluations to increase arbitrary quota Only switch to a model with a larger context window without rate-limit planning Rotate every customer account ID Explanation Correct Answer A:Provisioned throughput and bounded demand determine sustainable inference rates. Incorrect Answers B:Safety reports are not the main rate-limit allocation mechanism. C:Context capacity does not grant more request or token throughput. D:Identity changes do not increase project inference quota.
Question 405 An assistant needs to give spoken answers in the same language detected from the user's audio. What state should persist? Use the initial greeting's language even after detecting a language switch Persist transcript text but not the detected audio locale Choose any available synthesis voice without locale compatibility checks Detected language/locale and selected synthesis voice for the session Explanation Correct Answer D:Consistent input and output locale supports intelligible multi-turn speech. Incorrect Answers A:Session language may change; stale configuration can produce mismatched synthesized speech. B:Translated or normalized text may not reliably identify the speech language for voice selection. C:An arbitrary voice can have a different language or pronunciation profile.
Question 406 A Content Understanding workflow must choose between simple targeted fields and richer evidence-driven extraction. What design principle applies? Assume every analyzer supports all preview functions automatically Treat any unverified complex reasoning as a guaranteed extracted fact Disable field schemas to avoid design decisions Choose a documented analyzer configuration or mode matched to task and API support Explanation Correct Answer D:Capability, model support and version constraints differ across extraction approaches. Incorrect Answers A:GA and preview feature availability can differ. B:Sophisticated models still require source grounding and validation. C:Without explicit outputs, the downstream contract becomes unclear.
Question 407 A service connection cannot read a Foundry project although it has Azure subscription Reader. What authorization issue is plausible? The app should use an unlimited anonymous token The service needs appropriate project/resource role or data-plane permission The user should change the model sampling temperature The developer should add a DNS TXT record named Owner Explanation Correct Answer B:General resource visibility does not automatically grant inference and project operations. Incorrect Answers A:Removing authentication is not a safe fix for denied permissions. C:Temperature cannot grant missing access rights. D:A TXT label cannot create Azure RBAC assignments.
Question 408 An accessibility app produces only 'image' as alt text for complex diagrams. What should change? Provide concise functional content and meaningful visible relationships Generate random labels to increase length Use the filename alone for all alt text Add decorative adjectives unrelated to the data Explanation Correct Answer A:Accessible descriptions should communicate the information the diagram conveys. Incorrect Answers B:Invented details can actively mislead screen-reader users. C:A filename rarely contains sufficient functional context. D:Extra style does not convey the diagram's relationships.
Question 409 An image model generates a photorealistic employee portrait for an advertisement. What responsible control matters? Treat realism as proof the person consented Present it as a genuine employee photo without verification Remove all provenance metadata to avoid questions Check disclosure, consent and approved identity/brand usage before release Explanation Correct Answer D:Realistic synthetic portraits can imply identities or endorsements without authorization. Incorrect Answers A:Photographic plausibility does not grant rights. B:Synthetic output should not automatically be described as actual photography. C:Obscuring origin can mislead the campaign audience.
Question 410 A private model endpoint works from one VNet but not a peered app VNet. What should be verified? Only the embedding count in a search index Endpoint DNS, effective routes, peering and network policy on the client path Only the registered application name Only whether the model returns correct arithmetic Explanation Correct Answer B:Private connectivity requires resolution and reachable permitted network hops. Incorrect Answers A:Vector inventory does not control endpoint access. C:Naming cannot replace DNS or route verification. D:Output correctness does not prove the blocked client's network path.