Practice Test #3

Speech recognition converts audio into text. The scenario is about request IP and user-agent metadata, so an audio model would not solve suspicious traffic detection. Trap: The video streaming context can distract you toward media AI, but the actual data is API traffic metadata.

Named entity recognition extracts entities such as people, places, or organizations from text. It does not compare request sources against baseline behavior. Trap: A user-agent is a string, but the task is behavioral detection, not entity extraction from natural language.

Fraud forecasting predicts future fraud volume or trends. The requirement is to score each current request source against a baseline and flag unusual behavior immediately. Trap: Security and fraud both involve risk, but forecasting is not per-request anomaly scoring.

Fine-tuning can adapt model behavior or reduce reliance on examples, but it adds training, evaluation, and governance overhead that is not justified for a once-daily workload with satisfactory quality. Trap: It sounds like a way to replace few-shot examples, but the scenario asks for the simplest cost reduction while keeping current performance.

Increasing prompt tokens would send more input to the model and usually raise cost, which conflicts with the monthly cost objective. Trap: More examples can sometimes improve quality, but quality is already satisfactory and the goal is cost reduction.

Provisioned Throughput is intended for workloads that need reserved model capacity and predictable throughput; a single scheduled invocation per day does not justify it. Trap: Provisioned capacity can sound like a pricing discount, but for very low-volume usage it is usually less cost-effective than on-demand inference.

AWS Config evaluates AWS resource configurations against rules and compliance requirements. It does not generate PySpark code or guide a team through building an AWS Glue job. Trap: It is an AWS governance service, but the scenario is about developer assistance rather than configuration auditing.

Amazon Personalize builds recommendation experiences from user and item interaction data. It is unrelated to creating ETL code for AWS Glue. Trap: It is an AI service, but it solves recommendation problems, not developer productivity problems.

Amazon Comprehend analyzes text for entities, sentiment, key phrases, and similar NLP insights. It does not provide AWS console coding guidance for Glue jobs. Trap: CSV processing can involve text, but the requirement is to author an ETL job, not perform NLP on the data.

Amazon Inspector finds software vulnerabilities and unintended network exposure. It does not calculate fairness metrics or provide SHAP explanations for loan recommendations. Trap: The words security and compliance sound relevant to banking, but the requirement is responsible AI explainability.

Amazon Macie helps discover and classify sensitive data in Amazon S3. It can support privacy governance, but it cannot explain predictions or measure bias across demographic groups. Trap: Protecting training data matters, but encryption and sensitive-data discovery do not satisfy the transparency requirement.

Amazon Rekognition is a computer vision service for images and video. Adding image labels is unrelated to text and structured loan features and does not produce SHAP attributions. Trap: More labeled data can improve some models, but the question asks for audit-ready explanation artifacts.

Retraining may reduce unsafe outputs, but it does not provide deterministic per-image blocking, does not implement the 90% threshold, and violates the requirement to avoid retraining the base model. Trap: Model improvement sounds proactive, but compliance asks for live enforcement on every output.

A one-time validation can support release readiness, but it cannot inspect future generated or retrieved images and cannot log each inference-time decision. Trap: Holdout testing is useful governance evidence, but it is not a runtime safety control.

User feedback is reactive and may improve the system later, but it does not stop unsafe images from being displayed immediately when the moderation confidence is high. Trap: Feedback loops help continuous improvement, but the policy requires immediate blocking and logging.

Auto scaling inference endpoints help handle chat volume and latency, but they are performance and availability features rather than direct HIPAA or SOC 2 compliance evidence. Trap: The p95 latency clue makes autoscaling feel relevant, but the question asks what can be demonstrated for compliance.

Cost optimization is important for operations, but it is not a primary HIPAA or SOC 2 control category for protecting healthcare data. Trap: Governance can include cost management, but compliance audits focus first on security, privacy, monitoring, and control evidence.

Loosely coupled microservices can improve maintainability and resilience, but the architecture style itself does not prove data protection or security monitoring. Trap: Resilient design helps availability, yet it is indirect compared with encryption and threat detection.

Switching models does not automatically give the assistant access to private hospital PDFs, nightly refreshes, or citations. Trap: A different foundation model may change answer quality or latency, but it does not solve the missing data grounding problem.

Lower temperature makes generation less random, but it cannot retrieve current facts from S3 or provide citations. Trap: It can reduce variability and sometimes reduce hallucination, but it is an inference setting, not a knowledge integration feature.

Model invocation logging helps with audit and troubleshooting, but it does not ingest S3 documents, refresh a knowledge corpus, or cite sources in responses. Trap: Logging is useful for governance, but the scenario is asking for factual grounding and retrieval.

RAG retrieves external context for generative AI answers; it does not calculate tabular fairness metrics such as demographic parity difference. Trap: RAG is a common AI architecture pattern, but the scenario is about dataset bias before training.

Production model monitoring applies after a model is deployed, while the question asks for checks before training any model. Trap: Monitoring is part of responsible ML, but it is the wrong lifecycle phase.

Model documentation records intended use, data, metrics, and limitations, but it does not directly run bias metrics on the dataset. Trap: Documentation supports governance, but the requirement asks for an analytical bias check.

AUC evaluates how well a classifier separates positive and negative classes across thresholds. It does not compare two free-form summaries for semantic similarity. Trap: AUC is a familiar accuracy metric, but it belongs to classification rather than summarization.

F1 balances precision and recall for labels, spans, or overlap-style comparisons, but the question specifically asks for semantic similarity to reference summaries. Trap: F1 sounds like a general accuracy score, but it is not the best listed metric for meaning-preserving abstractive summaries.

Real world knowledge scoring targets factual knowledge or grounding behavior, not semantic similarity to a supplied reference summary. Trap: Factuality matters for summaries, but the question names semantic similarity to references as the evaluation goal.