AI-103 Premium File
- 187 Questions & Answers
- Last Update: Nov 19, 2026
The project team needs production acceptance evidence. Its application has this need: A product-search app must retrieve clothing images similar to a customer's uploaded photograph. The required outcome is to use visual similarity while keeping product metadata and permissions attached. Select TWO checks that test this outcome.
Check that removed or restricted products are absent from search results exercises a relevant negative or edge-case condition. It matters here because similar-looking products may differ in fit or material that images cannot establish. The complementary positive check—upload a known product image and verify returned IDs map to their catalog records—then confirms the intended business capability still works; neither observation alone settles the whole requirement.
Upload a known product image and verify returned IDs map to their catalog records measures the actual outcome needed to use visual similarity while keeping product metadata and permissions attached. It would expose the specific failure where visually similar recommendations display the wrong product's price and inventory. Pair it with check that removed or restricted products are absent from search results to check both successful behavior and the remaining operational boundary.
Retrieval success without catalog binding can surface irrelevant items. For this workload, the decisive functional evidence is: Upload a known product image and verify returned IDs map to their catalog records. The shortcut in this option could pass even when visually similar recommendations display the wrong product's price and inventory.
Revoked products must be removed from effective retrieval, not merely database views. The deployment needs to use visual similarity while keeping product metadata and permissions attached. An independent test is: Check that removed or restricted products are absent from search results.
The negative inventory-control path is untested. The real incident to guard against is that visually similar recommendations display the wrong product's price and inventory. This does not replace the case-specific validation: Check that removed or restricted products are absent from search results.
Consider an application deployed with Microsoft Foundry: A safety officer needs a structured inspection of image hazards and a concise evidence description. The owner wants to return a bounded JSON-style hazard report with visible support. Which implementation is most appropriate?
A typed report can separate detected evidence from uncertain risk labels; for this workload the decision is to use multimodal analysis plus a validated structured output schema and human escalation for ambiguous hazards. The acceptance result is: Test clear, obscured, and absent-object images against annotated reports. An independent edge-case check is: Verify uncertainty fields trigger reviewer escalation instead of automatic clearance. That distinction matters because visual hazard detection supports screening but does not substitute for an on-site inspection.
Missing evidence is not proof of safety, which does not satisfy the requirement to return a bounded JSON-style hazard report with visible support. A suitable test of the correct implementation is: Test clear, obscured, and absent-object images against annotated reports. This catches the particular incident where the report marks a partially obscured extinguisher as absent with full confidence.
A downstream inspection workflow may lose required fields. This shortcut can leave the following observable failure unresolved: The report marks a partially obscured extinguisher as absent with full confidence. A proper deployment instead must return a bounded JSON-style hazard report with visible support; the missing capability would leave the deployment exposed to the report marks a partially obscured extinguisher as absent with full confidence. The right approach is to use multimodal analysis plus a validated structured output schema and human escalation for ambiguous hazards, rather than relying on this unrelated shortcut. The distinguishing test is: Test clear, obscured, and absent-object images against annotated reports.
Policy conclusions can need additional context, so it cannot establish return a bounded JSON-style hazard report with visible support. The working design instead needs to use multimodal analysis plus a validated structured output schema and human escalation for ambiguous hazards. A strong boundary test is: Verify uncertainty fields trigger reviewer escalation instead of automatic clearance. This matters because visual hazard detection supports screening but does not substitute for an on-site inspection.
After rollout, the team observes: The report marks a partially obscured extinguisher as absent with full confidence. The intended design was to return a bounded JSON-style hazard report with visible support. Which root cause should be investigated first?
The failure that the report marks a partially obscured extinguisher as absent with full confidence points to the model failed to represent occlusion uncertainty in the output contract. The direct functional check is: Test clear, obscured, and absent-object images against annotated reports. Then test the independent boundary—verify uncertainty fields trigger reviewer escalation instead of automatic clearance—to distinguish a genuine repair from a passing routine response.
If the model mixed image references or generalized a feature across both inputs, then the expected symptom is the comparison says both rooms have balconies when only one image shows a balcony; this incident instead reports that the report marks a partially obscured extinguisher as absent with full confidence. The nearer diagnosis is the model failed to represent occlusion uncertainty in the output contract; to distinguish it from the alternative of the model mixed image references or generalized a feature across both inputs, test the specific service behavior: Test clear, obscured, and absent-object images against annotated reports.
If generation inferred an unsupported identity instead of describing observable features, then the expected symptom is alt text identifies a person who is not named anywhere in the image; this incident instead reports that the report marks a partially obscured extinguisher as absent with full confidence. The nearer diagnosis is the model failed to represent occlusion uncertainty in the output contract; to distinguish it from the alternative of generation inferred an unsupported identity instead of describing observable features, test the specific service behavior: Test clear, obscured, and absent-object images against annotated reports.
If the localization stage mismapped region coordinates or ignored the detected pixels, then the expected symptom is the model reports the correct defect category but draws boxes on clean metal; this incident instead reports that the report marks a partially obscured extinguisher as absent with full confidence. The nearer diagnosis is the model failed to represent occlusion uncertainty in the output contract; to distinguish it from the alternative of the localization stage mismapped region coordinates or ignored the detected pixels, test the specific service behavior: Test clear, obscured, and absent-object images against annotated reports.
A release reviewer considers this scenario: A safety officer needs a structured inspection of image hazards and a concise evidence description. The objective is to return a bounded JSON-style hazard report with visible support. Which TWO validation outcomes directly establish that behavior? Select TWO answers.
Test clear, obscured, and absent-object images against annotated reports measures the actual outcome needed to return a bounded JSON-style hazard report with visible support. It would expose the specific failure where the report marks a partially obscured extinguisher as absent with full confidence. Pair it with verify uncertainty fields trigger reviewer escalation instead of automatic clearance to check both successful behavior and the remaining operational boundary.
Verify uncertainty fields trigger reviewer escalation instead of automatic clearance exercises a relevant negative or edge-case condition. It matters here because visual hazard detection supports screening but does not substitute for an on-site inspection. The complementary positive check—test clear, obscured, and absent-object images against annotated reports—then confirms the intended business capability still works; neither observation alone settles the whole requirement.
Uncertainty and absence require separate annotated tests. For this workload, the decisive functional evidence is: Test clear, obscured, and absent-object images against annotated reports. The shortcut in this option could pass even when the report marks a partially obscured extinguisher as absent with full confidence.
No detection under occlusion does not demonstrate absence. The deployment needs to return a bounded JSON-style hazard report with visible support. An independent test is: Verify uncertainty fields trigger reviewer escalation instead of automatic clearance.
The review-required state must control downstream decisions, not just telemetry. The real incident to guard against is that the report marks a partially obscured extinguisher as absent with full confidence. This does not replace the case-specific validation: Verify uncertainty fields trigger reviewer escalation instead of automatic clearance.
The team has implemented the selected Azure design for this scenario: A safety officer needs a structured inspection of image hazards and a concise evidence description. It intends to return a bounded JSON-style hazard report with visible support. Which remaining operational limitation is specifically relevant?
Even after the team chooses to use multimodal analysis plus a validated structured output schema and human escalation for ambiguous hazards, visual hazard detection supports screening but does not substitute for an on-site inspection. This matters for a safety officer needs a structured inspection of image hazards and a concise evidence description. The edge-case check is: Verify uncertainty fields trigger reviewer escalation instead of automatic clearance. The regular success check alone—test clear, obscured, and absent-object images against annotated reports—cannot establish that this risk is controlled.
Watermarks can be cropped or altered after delivery, so downstream provenance needs controls could be important if a brand's campaign imagery requires visible ownership labels and consistent permitted symbols, but the actual use case is a safety officer needs a structured inspection of image hazards and a concise evidence description. The closest operational risk is visual hazard detection supports screening but does not substitute for an on-site inspection; evaluate that instead of watermarks can be cropped or altered after delivery, so downstream provenance needs controls by testing: Verify uncertainty fields trigger reviewer escalation instead of automatic clearance.
Image moderation can misclassify art, news, or medical contexts and needs review paths could be important if a photo-generation service for students must block disallowed imagery and preserve benign educational illustrations, but the actual use case is a safety officer needs a structured inspection of image hazards and a concise evidence description. The closest operational risk is visual hazard detection supports screening but does not substitute for an on-site inspection; evaluate that instead of image moderation can misclassify art, news, or medical contexts and needs review paths by testing: Verify uncertainty fields trigger reviewer escalation instead of automatic clearance.
Image-generation models may not reliably reproduce small exact lettering without downstream graphic design could be important if a design agency needs product concept images for a catalog, with exact aspect ratio and approved visual style, but the actual use case is a safety officer needs a structured inspection of image hazards and a concise evidence description. The closest operational risk is visual hazard detection supports screening but does not substitute for an on-site inspection; evaluate that instead of image-generation models may not reliably reproduce small exact lettering without downstream graphic design by testing: Verify uncertainty fields trigger reviewer escalation instead of automatic clearance.
An engineering review identifies this need: A public image-captioning agent may encounter screenshots containing hidden hostile instructions. To prevent image-borne text from controlling privileged tools, which configuration should the team choose?
Prompt injection can arrive through pixels as well as documents; for this workload the decision is to treat OCR and image text as untrusted data and enforce tool policies outside multimodal prompts. The acceptance result is: Embed a fake system message in the image and verify no data-transfer tool call occurs. An independent edge-case check is: Ask a legitimate visual question about the image and confirm it still answers. That distinction matters because visual prompt-shield coverage can be incomplete when text is stylized or occluded.
External images are not trusted operator messages, which does not satisfy the requirement to prevent image-borne text from controlling privileged tools. A suitable test of the correct implementation is: Embed a fake system message in the image and verify no data-transfer tool call occurs. This catches the particular incident where a screenshot instructs the assistant to send an internal report to an outside server.
Visual content cannot authorize data transfers. This shortcut can leave the following observable failure unresolved: A screenshot instructs the assistant to send an internal report to an outside server. A proper deployment instead must prevent image-borne text from controlling privileged tools; the missing capability would leave the deployment exposed to a screenshot instructs the assistant to send an internal report to an outside server. The right approach is to treat OCR and image text as untrusted data and enforce tool policies outside multimodal prompts, rather than relying on this unrelated shortcut. The distinguishing test is: Embed a fake system message in the image and verify no data-transfer tool call occurs.
Images can still contain malicious text and content, so it cannot establish prevent image-borne text from controlling privileged tools. The working design instead needs to treat OCR and image text as untrusted data and enforce tool policies outside multimodal prompts. A strong boundary test is: Ask a legitimate visual question about the image and confirm it still answers. This matters because visual prompt-shield coverage can be incomplete when text is stylized or occluded.
A Foundry application has a specific incident: A screenshot instructs the assistant to send an internal report to an outside server. The operator is trying to prevent image-borne text from controlling privileged tools. Which failure mechanism best explains the evidence?
The failure that a screenshot instructs the assistant to send an internal report to an outside server points to image-extracted instructions were promoted into agent tool authority. The direct functional check is: Embed a fake system message in the image and verify no data-transfer tool call occurs. Then test the independent boundary—ask a legitimate visual question about the image and confirm it still answers—to distinguish a genuine repair from a passing routine response.
If the visual reasoning pipeline ignored the occluded exit marking or used unrelated evidence, then the expected symptom is the agent claims the exit is clear despite the pallet covering the sign; this incident instead reports that a screenshot instructs the assistant to send an internal report to an outside server. The nearer diagnosis is image-extracted instructions were promoted into agent tool authority; to distinguish it from the alternative of the visual reasoning pipeline ignored the occluded exit marking or used unrelated evidence, test the specific service behavior: Embed a fake system message in the image and verify no data-transfer tool call occurs.
If generation inferred an unsupported identity instead of describing observable features, then the expected symptom is alt text identifies a person who is not named anywhere in the image; this incident instead reports that a screenshot instructs the assistant to send an internal report to an outside server. The nearer diagnosis is image-extracted instructions were promoted into agent tool authority; to distinguish it from the alternative of generation inferred an unsupported identity instead of describing observable features, test the specific service behavior: Embed a fake system message in the image and verify no data-transfer tool call occurs.
If the edit did not preserve temporal alignment at the clip boundary, then the expected symptom is the replacement slate flickers and the next scene starts mid-word; this incident instead reports that a screenshot instructs the assistant to send an internal report to an outside server. The nearer diagnosis is image-extracted instructions were promoted into agent tool authority; to distinguish it from the alternative of the edit did not preserve temporal alignment at the clip boundary, test the specific service behavior: Embed a fake system message in the image and verify no data-transfer tool call occurs.
A reviewer must verify that the solution will prevent image-borne text from controlling privileged tools. The deployment context is: A public image-captioning agent may encounter screenshots containing hidden hostile instructions. Which TWO tests provide the strongest evidence? Select TWO answers.
Embed a fake system message in the image and verify no data-transfer tool call occurs measures the actual outcome needed to prevent image-borne text from controlling privileged tools. It would expose the specific failure where a screenshot instructs the assistant to send an internal report to an outside server. Pair it with ask a legitimate visual question about the image and confirm it still answers to check both successful behavior and the remaining operational boundary.
Ask a legitimate visual question about the image and confirm it still answers exercises a relevant negative or edge-case condition. It matters here because visual prompt-shield coverage can be incomplete when text is stylized or occluded. The complementary positive check—embed a fake system message in the image and verify no data-transfer tool call occurs—then confirms the intended business capability still works; neither observation alone settles the whole requirement.
Recognizing malicious text is not the same as preventing its execution. For this workload, the decisive functional evidence is: Embed a fake system message in the image and verify no data-transfer tool call occurs. The shortcut in this option could pass even when a screenshot instructs the assistant to send an internal report to an outside server.
The system should still answer legitimate visual questions without following embedded commands. The deployment needs to prevent image-borne text from controlling privileged tools. An independent test is: Ask a legitimate visual question about the image and confirm it still answers.
Image pixels are untrusted content, not a source of tool authorization. The real incident to guard against is that a screenshot instructs the assistant to send an internal report to an outside server. This does not replace the case-specific validation: Ask a legitimate visual question about the image and confirm it still answers.
A team implementing Microsoft Foundry faces this scenario: A photo-generation service for students must block disallowed imagery and preserve benign educational illustrations. The nonnegotiable requirement is to apply multimodal safety policies at request and output boundaries. Which design choice best meets it?
Safety filters should cover input intent and unexpected output content; for this workload the decision is to configure supported moderation and evaluate both prompts and generated images. The acceptance result is: Test disallowed and benign prompt categories and inspect final images. An independent edge-case check is: Check that reasonable scientific diagrams are not overblocked. That distinction matters because image moderation can misclassify art, news, or medical contexts and needs review paths.
Watermarks address attribution, not policy classification, which does not satisfy the requirement to apply multimodal safety policies at request and output boundaries. A suitable test of the correct implementation is: Test disallowed and benign prompt categories and inspect final images. This catches the particular incident where a benign classroom diagram request yields an unsafe graphic.
Generated content can deviate from permitted intent. This shortcut can leave the following observable failure unresolved: A benign classroom diagram request yields an unsafe graphic. A proper deployment instead must apply multimodal safety policies at request and output boundaries; the missing capability would leave the deployment exposed to a benign classroom diagram request yields an unsafe graphic. The right approach is to configure supported moderation and evaluate both prompts and generated images, rather than relying on this unrelated shortcut. The distinguishing test is: Test disallowed and benign prompt categories and inspect final images.
A claimed educational purpose is not a universal exception, so it cannot establish apply multimodal safety policies at request and output boundaries. The working design instead needs to configure supported moderation and evaluate both prompts and generated images. A strong boundary test is: Check that reasonable scientific diagrams are not overblocked. This matters because image moderation can misclassify art, news, or medical contexts and needs review paths.
During incident response, telemetry shows: A benign classroom diagram request yields an unsafe graphic. The feature exists to apply multimodal safety policies at request and output boundaries. What is the most relevant diagnosis?
The failure that a benign classroom diagram request yields an unsafe graphic points to output moderation was omitted or thresholds were misconfigured. The direct functional check is: Test disallowed and benign prompt categories and inspect final images. Then test the independent boundary—check that reasonable scientific diagrams are not overblocked—to distinguish a genuine repair from a passing routine response.
If the model mixed image references or generalized a feature across both inputs, then the expected symptom is the comparison says both rooms have balconies when only one image shows a balcony; this incident instead reports that a benign classroom diagram request yields an unsafe graphic. The nearer diagnosis is output moderation was omitted or thresholds were misconfigured; to distinguish it from the alternative of the model mixed image references or generalized a feature across both inputs, test the specific service behavior: Test disallowed and benign prompt categories and inspect final images.
If the visual reasoning pipeline ignored the occluded exit marking or used unrelated evidence, then the expected symptom is the agent claims the exit is clear despite the pallet covering the sign; this incident instead reports that a benign classroom diagram request yields an unsafe graphic. The nearer diagnosis is output moderation was omitted or thresholds were misconfigured; to distinguish it from the alternative of the visual reasoning pipeline ignored the occluded exit marking or used unrelated evidence, test the specific service behavior: Test disallowed and benign prompt categories and inspect final images.
If generation inferred an unsupported identity instead of describing observable features, then the expected symptom is alt text identifies a person who is not named anywhere in the image; this incident instead reports that a benign classroom diagram request yields an unsafe graphic. The nearer diagnosis is output moderation was omitted or thresholds were misconfigured; to distinguish it from the alternative of generation inferred an unsupported identity instead of describing observable features, test the specific service behavior: Test disallowed and benign prompt categories and inspect final images.
The project team needs production acceptance evidence. Its application has this need: A photo-generation service for students must block disallowed imagery and preserve benign educational illustrations. The required outcome is to apply multimodal safety policies at request and output boundaries. Select TWO checks that test this outcome.
Test disallowed and benign prompt categories and inspect final images measures the actual outcome needed to apply multimodal safety policies at request and output boundaries. It would expose the specific failure where a benign classroom diagram request yields an unsafe graphic. Pair it with check that reasonable scientific diagrams are not overblocked to check both successful behavior and the remaining operational boundary.
Check that reasonable scientific diagrams are not overblocked exercises a relevant negative or edge-case condition. It matters here because image moderation can misclassify art, news, or medical contexts and needs review paths. The complementary positive check—test disallowed and benign prompt categories and inspect final images—then confirms the intended business capability still works; neither observation alone settles the whole requirement.
Overblocking benign scientific diagrams is also a safety-control failure. For this workload, the decisive functional evidence is: Test disallowed and benign prompt categories and inspect final images. The shortcut in this option could pass even when a benign classroom diagram request yields an unsafe graphic.
The rendered output can contain disallowed content despite a benign prompt classification. The deployment needs to apply multimodal safety policies at request and output boundaries. An independent test is: Check that reasonable scientific diagrams are not overblocked.
Allowed-path success cannot establish rejection of prohibited outputs. The real incident to guard against is that a benign classroom diagram request yields an unsafe graphic. This does not replace the case-specific validation: Check that reasonable scientific diagrams are not overblocked.
Please check your mailbox for a message from support@examlabs.com and follow the directions.