AI-103 Premium File
- 187 Questions & Answers
- Last Update: Nov 19, 2026
The project team needs production acceptance evidence. Its application has this need: An industrial inspection app needs labeled locations of visible rust rather than a generic photo caption. The required outcome is to provide region-level outputs tied to image coordinates. Select TWO checks that test this outcome.
Compare predicted boxes to human-annotated defect regions measures the actual outcome needed to provide region-level outputs tied to image coordinates. It would expose the specific failure where the model reports the correct defect category but draws boxes on clean metal. Pair it with test an image with no rust and verify no fabricated boxes are emitted to check both successful behavior and the remaining operational boundary.
Test an image with no rust and verify no fabricated boxes are emitted exercises a relevant negative or edge-case condition. It matters here because localization may miss fine corrosion where resolution is low or textures are ambiguous. The complementary positive check—compare predicted boxes to human-annotated defect regions—then confirms the intended business capability still works; neither observation alone settles the whole requirement.
Correct counts can still represent entirely wrong regions. For this workload, the decisive functional evidence is: Compare predicted boxes to human-annotated defect regions. The shortcut in this option could pass even when the model reports the correct defect category but draws boxes on clean metal.
False-positive behavior on absent defects must be measured. The deployment needs to provide region-level outputs tied to image coordinates. An independent test is: Test an image with no rust and verify no fabricated boxes are emitted.
Overbroad boxes fail localized defect detection quality. The real incident to guard against is that the model reports the correct defect category but draws boxes on clean metal. This does not replace the case-specific validation: Test an image with no rust and verify no fabricated boxes are emitted.
A wind-farm maintenance team has baseline and current drone photographs of turbine blades. The team must identify new corrosion only when the same blade section is compared across dates. Which multimodal design best meets this requirement?
The application must first establish that both images represent the same turbine blade and section. Registering comparable regions makes change detection meaningful; the model can then cite visible before/after evidence and flag uncertainty where angles or occlusion prevent a supported conclusion. The validation should include paired images of the same labeled blade under controlled viewpoint changes, plus a deliberately mismatched pair that must not generate a confident trend.
A single photograph may reveal corrosion, but without a matched baseline it cannot demonstrate that a particular defect is new. Historical visual evidence is a necessary part of the requested change decision.
Upload order does not align the same blade section or compensate for different viewpoints and lighting. Pixel differences from camera angle or shadows are not independently evidence of physical corrosion growth.
A maintenance ticket is useful context but cannot replace analysis of the actual image pair. It might describe prior damage, yet it cannot substantiate a visually detected change at a specific blade location.
A turbine inspection assistant reports a new crack using two drone images. Metadata then shows the baseline image is from blade A, while the current image is from blade B; no same-component comparison was performed. Which failure mechanism best explains the unsupported change claim?
Images of different blades cannot support a claim that a crack newly appeared on one blade. The application must bind images to stable blade and section identifiers and validate a comparable view before issuing a temporal change judgment. This explains the specific false temporal claim without requiring any failure of the underlying visual classifier on either photograph viewed independently.
Missing ticket provenance could reduce auditability, but it does not explain why the evidence compares photographs of two different physical blades. The wrong pair was selected before the report was composed.
A confidence threshold can affect alert frequency, but the provided metadata identifies a more direct defect: the baseline and current observations refer to different components, regardless of the score cutoff. The remedy is to reject invalid cross-component comparisons before the scoring stage, then evaluate thresholds only on genuine same-section image pairs.
Timezone rendering can move a displayed date, not change the subject identified in each image. The central failure is the missing image-pair identity and alignment validation. Correcting the report timestamp would not make photographs of blade A and blade B a valid before-and-after comparison.
A turbine-inspection application compares baseline and current drone imagery to identify newly visible damage on the same blade section. Which TWO acceptance tests directly demonstrate trustworthy visual change detection? Select TWO answers.
An annotated positive example measures whether the change detector finds a known new defect at the correct physical location after comparable regions are aligned. This verifies both the visual inference and its spatial grounding. Use the same source blade in both photographs and retain the labeled defect coordinates so the reported changed region can be checked independently.
The negative test checks that the system enforces the required same-blade identity boundary. A mismatched image pair must be rejected or explicitly labeled inconclusive rather than be treated as evidence of new corrosion. A reliable implementation should also identify why the pair was invalid, enabling an inspector to recapture comparable evidence instead of trusting a spurious trend.
A current-only inspection can identify visible conditions, but it lacks a baseline for the requirement to distinguish newly developed defects from pre-existing conditions. A newly visible crack and an old crack can look identical in the current image, so the single-frame check cannot establish change across inspection dates.
Differences between unrelated blades are not temporal evidence about either component. A reliable workflow must refuse this pairing rather than produce a confident new-damage finding. The system should compare stable blade IDs and section metadata first; successful rejection is the meaningful negative-case outcome, not a high image-difference score.
Matching JPEG or PNG formats provides no assurance that the photographs cover the same blade location. Component identity, viewpoint and visual evidence are the relevant test properties. The negative case must show the identity and alignment checks work, because matching file encodings cannot rule out two unrelated blade sections.
After launching a drone-image comparison workflow for turbine blades, which operational limitation remains important even if the change detector works on clean baseline/current examples?
Temporal change detection is sensitive to image registration, viewpoint, lighting, and stable component identity. Teams should preserve capture metadata, reject unmatched pairs and spot-check difficult locations before claiming a newly developed defect. A small shift in camera angle can expose a previously hidden surface and create a false positive, while heavy shadows can hide a real crack; paired-image review must cover both cases.
A watermark identifies media ownership; it does not validate the visual inference or confirm the same blade section was compared. The actual risk concerns comparability of baseline and current evidence.
A fluent model description is not an authoritative component identifier. Identity and image-registration checks must use source metadata and observable features, not narrative confidence alone. The model could confidently call both images turbine blades while still comparing photographs of different serial-numbered components; metadata remains the decisive identity check.
Confidence scores on independent images cannot establish physical growth when the source images do not refer to the same blade section. Paired identity and region comparability are prerequisites. The operator must establish the physical correspondence of the pixels before attributing any detected difference to new deterioration.
A team implementing Microsoft Foundry faces this scenario: A retail visual assistant must answer questions about a store shelf from one high-resolution image. The nonnegotiable requirement is to choose a vision analysis path that can zoom or crop as needed for small labels. Which design choice best meets it?
Fine-grained visual evidence may require higher-resolution regions than whole-image captions; for this workload the decision is to use image-aware reasoning with targeted crops for small text and verify findings against pixels. The acceptance result is: Compare extracted label values with a high-resolution manually verified crop. An independent edge-case check is: Mark unreadable labels as uncertain instead of supplying plausible numbers. That distinction matters because even cropped images may be affected by glare and perspective distortion.
Embeddings do not directly read digits, which does not satisfy the requirement to choose a vision analysis path that can zoom or crop as needed for small labels. A suitable test of the correct implementation is: Compare extracted label values with a high-resolution manually verified crop. This catches the particular incident where the assistant reads 8.99 as 3.99 from a tiny price label.
Statistical ranges are not evidence for this shelf. This shortcut can leave the following observable failure unresolved: The assistant reads 8.99 as 3.99 from a tiny price label. A proper deployment instead must choose a vision analysis path that can zoom or crop as needed for small labels; the missing capability would leave the deployment exposed to the assistant reads 8.99 as 3.99 from a tiny price label. The right approach is to use image-aware reasoning with targeted crops for small text and verify findings against pixels, rather than relying on this unrelated shortcut. The distinguishing test is: Compare extracted label values with a high-resolution manually verified crop.
Resizing can erase decisive characters, so it cannot establish choose a vision analysis path that can zoom or crop as needed for small labels. The working design instead needs to use image-aware reasoning with targeted crops for small text and verify findings against pixels. A strong boundary test is: Mark unreadable labels as uncertain instead of supplying plausible numbers. This matters because even cropped images may be affected by glare and perspective distortion.
A Foundry application has a specific incident: The assistant reads 8.99 as 3.99 from a tiny price label. The operator is trying to choose a vision analysis path that can zoom or crop as needed for small labels. Which failure mechanism best explains the evidence?
The failure that the assistant reads 8.99 as 3.99 from a tiny price label points to insufficient pixel resolution or crop selection corrupted text recognition. The direct functional check is: Compare extracted label values with a high-resolution manually verified crop. Then test the independent boundary—mark unreadable labels as uncertain instead of supplying plausible numbers—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 the assistant reads 8.99 as 3.99 from a tiny price label. The nearer diagnosis is insufficient pixel resolution or crop selection corrupted text recognition; 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: Compare extracted label values with a high-resolution manually verified crop.
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 assistant reads 8.99 as 3.99 from a tiny price label. The nearer diagnosis is insufficient pixel resolution or crop selection corrupted text recognition; to distinguish it from the alternative of generation inferred an unsupported identity instead of describing observable features, test the specific service behavior: Compare extracted label values with a high-resolution manually verified crop.
If the model failed to represent occlusion uncertainty in the output contract, then the expected symptom is the report marks a partially obscured extinguisher as absent with full confidence; this incident instead reports that the assistant reads 8.99 as 3.99 from a tiny price label. The nearer diagnosis is insufficient pixel resolution or crop selection corrupted text recognition; to distinguish it from the alternative of the model failed to represent occlusion uncertainty in the output contract, test the specific service behavior: Compare extracted label values with a high-resolution manually verified crop.
A reviewer must verify that the solution will choose a vision analysis path that can zoom or crop as needed for small labels. The deployment context is: A retail visual assistant must answer questions about a store shelf from one high-resolution image. Which TWO tests provide the strongest evidence? Select TWO answers.
Compare extracted label values with a high-resolution manually verified crop measures the actual outcome needed to choose a vision analysis path that can zoom or crop as needed for small labels. It would expose the specific failure where the assistant reads 8.99 as 3.99 from a tiny price label. Pair it with mark unreadable labels as uncertain instead of supplying plausible numbers to check both successful behavior and the remaining operational boundary.
Mark unreadable labels as uncertain instead of supplying plausible numbers exercises a relevant negative or edge-case condition. It matters here because even cropped images may be affected by glare and perspective distortion. The complementary positive check—compare extracted label values with a high-resolution manually verified crop—then confirms the intended business capability still works; neither observation alone settles the whole requirement.
An unreadable preview is not dependable evidence for numeric label extraction. For this workload, the decisive functional evidence is: Compare extracted label values with a high-resolution manually verified crop. The shortcut in this option could pass even when the assistant reads 8.99 as 3.99 from a tiny price label.
Invented numeric identifiers create downstream inventory errors. The deployment needs to choose a vision analysis path that can zoom or crop as needed for small labels. An independent test is: Mark unreadable labels as uncertain instead of supplying plausible numbers.
Nonempty output can be a fabricated or transposed identifier. The real incident to guard against is that the assistant reads 8.99 as 3.99 from a tiny price label. This does not replace the case-specific validation: Mark unreadable labels as uncertain instead of supplying plausible numbers.
An Azure AI engineer must address the following production request: A product-search app must retrieve clothing images similar to a customer's uploaded photograph. The solution must use visual similarity while keeping product metadata and permissions attached. What should be implemented?
Image embeddings enable similarity search but results need verified catalog metadata; for this workload the decision is to create or query an appropriate multimodal/vector index of authorized product images and identifiers. The acceptance result is: Upload a known product image and verify returned IDs map to their catalog records. An independent edge-case check is: Check that removed or restricted products are absent from search results. That distinction matters because similar-looking products may differ in fit or material that images cannot establish.
Sentiment does not encode garment appearance, which does not satisfy the requirement to use visual similarity while keeping product metadata and permissions attached. A suitable test of the correct implementation is: Upload a known product image and verify returned IDs map to their catalog records. This catches the particular incident where visually similar recommendations display the wrong product's price and inventory.
Unstructured manual descriptions can miss visual similarity signals. This shortcut can leave the following observable failure unresolved: Visually similar recommendations display the wrong product's price and inventory. A proper deployment instead must use visual similarity while keeping product metadata and permissions attached; the missing capability would leave the deployment exposed to visually similar recommendations display the wrong product's price and inventory. The right approach is to create or query an appropriate multimodal/vector index of authorized product images and identifiers, rather than relying on this unrelated shortcut. The distinguishing test is: Upload a known product image and verify returned IDs map to their catalog records.
Similarity score alone does not provide stock or item ID, so it cannot establish use visual similarity while keeping product metadata and permissions attached. The working design instead needs to create or query an appropriate multimodal/vector index of authorized product images and identifiers. A strong boundary test is: Check that removed or restricted products are absent from search results. This matters because similar-looking products may differ in fit or material that images cannot establish.
During incident response, telemetry shows: Visually similar recommendations display the wrong product's price and inventory. The feature exists to use visual similarity while keeping product metadata and permissions attached. What is the most relevant diagnosis?
The failure that visually similar recommendations display the wrong product's price and inventory points to embedding hits were not joined correctly to authoritative catalog identifiers. The direct functional check is: Upload a known product image and verify returned IDs map to their catalog records. Then test the independent boundary—check that removed or restricted products are absent from search results—to distinguish a genuine repair from a passing routine response.
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 visually similar recommendations display the wrong product's price and inventory. The nearer diagnosis is embedding hits were not joined correctly to authoritative catalog identifiers; to distinguish it from the alternative of generation inferred an unsupported identity instead of describing observable features, test the specific service behavior: Upload a known product image and verify returned IDs map to their catalog records.
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 visually similar recommendations display the wrong product's price and inventory. The nearer diagnosis is embedding hits were not joined correctly to authoritative catalog identifiers; 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: Upload a known product image and verify returned IDs map to their catalog records.
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 visually similar recommendations display the wrong product's price and inventory. The nearer diagnosis is embedding hits were not joined correctly to authoritative catalog identifiers; to distinguish it from the alternative of the model mixed image references or generalized a feature across both inputs, test the specific service behavior: Upload a known product image and verify returned IDs map to their catalog records.
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.
Please check your mailbox for a message from support@examlabs.com and follow the directions.