Results

Explanation

RAG retrieves the relevant passages from a current document store at query time and passes them to the model, so answers reflect the latest version and can cite their source. Amazon Bedrock Knowledge Bases manages the ingestion, chunking, embedding, and retrieval, and re-syncing after a policy change is a data update, not a training job. Option A retrains monthly at real cost and still cannot cite a source. Option C is the most expensive customisation method and is meant for teaching a model a domain's language, not for keeping facts current. Option D breaks on context-window limits as the policy set grows and multiplies token cost on every request.

Explanation

Every model has a context window: the maximum number of tokens it can take in as input plus produce as output in one request. A 400-page document exceeds it, so the request fails or the model only attends to what fits. The fix is to chunk the document and summarise in stages, or to use retrieval so that only the relevant passages are sent. Option A confuses temperature, which controls randomness of the output, with input capacity. Option C would not change the context limit; a fine-tuned model has the same window. Option D refers to embeddings, which are used for semantic search and retrieval, not for how much text a model can read at once.

Explanation

A ranking model that learns patterns from historical data is classic machine learning: it predicts, it does not create. A system that produces new text from a brief is generative AI. A system that pursues a goal by planning multiple steps, calling tools or APIs, and adjusting to intermediate results is agentic AI, a term the 2026 exam guide added alongside AI, ML, and GenAI. Option A labels the recommender "deep learning", which is a technique (neural networks with many layers), not a description of what the system does, and it drops agentic AI entirely. Options C and D shuffle the labels so that the API-calling workflow is called generative or the recommender is called agentic, neither of which fits.

Explanation

A model learns the patterns in its training data, and a decade of human hiring decisions can encode discrimination that the model then reproduces and amplifies. The exam guide expects you to recognise bias, fairness, and inclusivity as responsible-AI features, to know dataset characteristics such as balance and representativeness, and to use tools such as subgroup analysis, label-quality checks, and human audits. SageMaker Clarify measures pre-training data bias and post-training prediction bias across groups. Option A confuses a sampling parameter with a fairness problem. Option C addresses model fit, not disparate outcomes. Option D exposes real candidates to harm and legal risk before the problem is understood.

Explanation

AWS PrivateLink provides interface VPC endpoints that place a private IP address for the Bedrock service inside your VPC, so requests travel over the AWS network and never touch the public internet; the exam guide names PrivateLink among the services that secure AI systems. Option A still sends traffic to a public endpoint, merely from a shared public IP, which violates the stated policy. Option C protects confidentiality in transit but the path is still the public internet, which is what the policy forbids. Option D connects an on-premises network to AWS and is neither necessary nor sufficient for a workload already inside a VPC.

Explanation

Fine-tuning adjusts the model's weights using labelled examples of the behaviour you want, and it is the right tool when the gap is consistent style, tone, or output format rather than facts. Thousands of prompt-and-response pairs are exactly the training set it needs. Option A adds knowledge, which the scenario says is not the problem. Option C changes how much input the model can read, not how it phrases or structures output. Option D concerns retrieval infrastructure and has no effect on style. The exam expects you to separate knowledge problems, solved by RAG, from behaviour problems, solved by fine-tuning.

Explanation

An embedding model converts text into a numeric vector that captures meaning, so "money-back guarantee" and "refund policy" end up close together in vector space even though they share no words. Storing document embeddings in a vector database and searching by similarity is what lets the query find the page; it is also the retrieval step inside RAG. Option A is exactly the approach that fails here, because keyword matching requires the words to appear. Option C is expensive and does not by itself provide search; a fine-tuned model still needs a retrieval mechanism to find documents. Option D describes tokenisation, a preprocessing step that has nothing to do with semantic matching.

Explanation

Batch inference processes a large dataset offline on a schedule, which is exactly the nightly 40-million-record job: no client is waiting, so throughput matters more than latency. Real-time inference serves individual requests synchronously with low latency, which is what a 300 ms checkout score demands. Option B wastes money keeping a low-latency endpoint warm for a job nobody waits for. Option C would make checkout wait for a scheduled run. Option D misuses both terms: asynchronous inference queues individual requests with large payloads and returns results later via notification, and batch is unsuitable for a per-transaction check. The exam guide also lists serverless inference, which auto-scales for intermittent traffic and is not asked for here.

Explanation

Continued pre-training extends a model's original self-supervised training on a large body of domain text, and it needs no labels, which matches the millions of unlabelled papers and the goal of absorbing an entire field's language. Option A requires labelled examples the company does not have and teaches specific task behaviour rather than broad domain fluency. Option C would let the model look up papers at query time but would not change how fluently it reads or writes the domain's language. Option D can nudge style within a single request but cannot teach the vocabulary of a whole discipline. Note that continued pre-training is also the most expensive of these approaches.

Explanation

Diffusion models generate images by iteratively refining random noise toward a picture that matches the prompt; Amazon Nova Canvas and Stability AI models in Amazon Bedrock are examples. A multimodal model accepts more than one kind of input, such as an image and text together, and can reason across them to answer a question. Option B assigns image generation to a text LLM, which produces text, and question answering to an embedding model, which produces vectors rather than answers. Option C uses embeddings for tasks that require generation. Option D swaps the two roles; a diffusion model generates images and does not answer questions about them.

Explanation

Amazon Bedrock Guardrails applies configurable safeguards to prompts and responses independently of the model: denied topics block whole subject areas such as investment advice, content filters catch categories like hate and violence, sensitive information filters detect and redact PII, word filters block specific terms, and a contextual grounding check flags unsupported claims. The exam guide names Guardrails as the tool for identifying responsible-AI features. Option A documents a model; it filters nothing at run time. Option C discovers sensitive data at rest in Amazon S3 and does not sit in the chat path. Option D detects bias and explains predictions; it is not a content filter.

Explanation

Under the shared responsibility model, AWS secures the cloud: the physical facilities, hardware, network, and managed service software, including the hosts that run foundation models. The customer secures what they put in the cloud: who may call the service through IAM, how data is classified and encrypted, what guardrails and prompts are applied, and how the application authenticates users and handles output. Option B assigns customer-side configuration to AWS. Option C assigns AWS-managed infrastructure to the customer; there are no model hosts for the customer to patch. Option D is the dangerous misconception the model exists to correct.

Explanation

Supervised learning trains on examples with known outcomes, and "repaid" or "defaulted" is exactly that label, so default prediction is supervised. Discovering groupings with no predefined categories is clustering, an unsupervised technique that finds structure in unlabelled data. Option B inverts the two. Option C proposes reinforcement learning, which learns from reward signals through trial and error, as in game playing or robotics, and has no place in either task. Option D fails on the marketing project: there are no labels to supervise with, which is the whole reason the team wants the segments discovered rather than assigned.

Explanation

Distillation transfers a large teacher model's behaviour on a task to a smaller student model, typically by generating high-quality responses from the teacher and fine-tuning the student on them. The result is cheaper and faster inference with accuracy close to the teacher's on that task, and Amazon Bedrock offers it as a managed customisation method; the 2026 exam guide lists distillation among the customisation cost trade-offs. Option B does not change speed or cost; temperature only alters randomness. Option C buys guaranteed capacity for the same expensive model and does not lower per-token cost or latency. Option D moves in the wrong direction on both cost and speed.

Explanation

Bedrock on-demand pricing is token-based: you pay for every input token sent and every output token generated. A long system prompt repeated on each call multiplies input cost, and unbounded answers multiply output cost, so trimming the prompt and capping output tokens attack both halves of the bill. Option C is wrong because temperature controls randomness, not length or speed, and has no effect on price. Option D is not a real pricing switch; on-demand inference already is real-time. Option E would allow even more tokens per request and is a model property, not a setting you dial to save money.

Explanation

The exam guide asks for the cost trade-offs of pre-training, fine-tuning, in-context learning, RAG, and distillation. Prompt engineering, a form of in-context learning, only changes the request, so it costs nothing beyond the tokens sent and should be tried first. Continued pre-training processes enormous amounts of text and sits at the expensive end. Option C is false: fine-tuning consumes training compute, needs labelled data, and the custom model then needs Provisioned Throughput or an on-demand custom-model deployment. Option D is the opposite of RAG's advantage; updating the document store needs no training. Option E reverses distillation, which produces a smaller model.

Explanation

A large gap between near-perfect training accuracy and poor accuracy on unseen data is the signature of overfitting: the model has memorised the training set, including its noise, instead of learning patterns that generalise. Standard remedies are more and more varied training data, regularisation, early stopping, or a simpler model. Option A describes the opposite problem; underfitting shows poor accuracy on training and test data alike, and adding capacity would worsen the memorisation. Option C addresses class imbalance, which is not indicated by the symptoms given. Option D confuses the terms: high bias is another name for underfitting, and simplifying an already overfit model is a valid move only if you diagnose it correctly as high variance.

Explanation

The exam guide lists intellectual property infringement claims, biased model outputs, loss of customer trust, end-user risk, and hallucinations as legal risks of generative AI. Copied protected content and confidently false statements are two of the clearest examples, and both can arrive without any malicious intent on the company's part. Option C is a cost-management issue, not a legal risk. Option D is an ordinary architecture detail with no legal dimension. Option E describes normal, expected operation; prompts are almost always shorter than the context window.

Explanation

Temperature controls how much randomness the model uses when choosing the next token. Near zero, the model almost always picks the most probable token, giving consistent, repeatable, factual output, which is what extraction needs. Higher values flatten the probability distribution so less likely tokens are chosen more often, producing the variety a brainstorm wants. Option A inverts the two. Option C is false: variability is precisely what temperature governs, along with related parameters such as top-p and top-k. Option D is wrong because the maximum token count only caps how long the answer can be; it does not change which words are chosen.

Explanation

Amazon Bedrock's data-handling commitments are frequently tested: your content is not used to improve the base models, is not shared with any third-party model provider, is encrypted in transit and at rest, and is processed and stored in the Region where you use the service, with the option of customer-managed KMS keys for encryption. This is also the answer to data-residency questions under governance. Option A describes the opposite of the commitment. Option C would break residency requirements that many regulated customers depend on. Option D invents a retention and review practice that does not exist; model invocation logging is opt-in and goes to the customer's own S3 bucket or CloudWatch Logs.

Explanation

A knowledge base implements the RAG pipeline: ingestion splits documents into chunks, an embedding model turns each chunk into a vector, and the vectors are written to a vector store such as Amazon OpenSearch Serverless, Aurora PostgreSQL, Neptune Analytics, or Amazon S3 Vectors. At query time the user's question is embedded, the most similar chunks are retrieved, and they are supplied to the model as context. Option A describes fine-tuning, which a knowledge base never does. Option C is impossible at this scale because of context-window limits and cost. Option D invents a summarisation step; Comprehend is a separate NLP service and is not part of knowledge-base ingestion.

Explanation

The exam guide lists the foundation model lifecycle as data selection, model selection, pre-training, fine-tuning, evaluation, deployment, and feedback. The logic is that you choose or gather data and pick a model architecture first, pre-train on broad data, adapt with fine-tuning, evaluate before releasing, deploy, and then use production feedback to drive the next iteration. The real exam tests sequences like this with ordering items. Options B, C, and D each place deployment or fine-tuning before pre-training or data selection, which cannot happen: a model cannot be deployed or adapted before it exists, and it cannot be trained before data has been chosen.

Explanation

Machine learning produces predictions with some error rate. When the correct answer is fully specified by known rules and must be exact, as with tax calculation, a deterministic implementation is cheaper, auditable, and correct every time, and the exam guide names this as a case where AI/ML is not appropriate. Option A would approximate rules the team already knows, introducing errors where none are necessary. Option C adds hallucination risk to a task that tolerates none, and a foundation model is not a calculator. Option D misapplies Amazon Personalize, which builds recommendation systems from user behaviour and has nothing to do with rule-based computation.

Explanation

The exam guide names Amazon OpenSearch Service, Amazon Aurora, Amazon Neptune, and Amazon RDS for PostgreSQL as AWS services that store embeddings for vector search, and Amazon Bedrock Knowledge Bases can create an OpenSearch Serverless collection or an Aurora PostgreSQL Serverless store for you; Amazon S3 Vectors is a newer option as well. Option C is a text-to-speech service. Option D records API activity for auditing and stores no application data. Option E analyses images and video. None of those three has a vector index or a similarity search capability.

Explanation

The exam guide asks you to identify trade-offs between model safety and transparency and to measure interpretability against performance. A model whose decisions must be explained to applicants and regulators needs either intrinsic interpretability, which simple models provide, or post-hoc explanations such as the SHAP feature attributions SageMaker Clarify produces for complex models. Either path is defensible if the trade-off is recorded. Option A ignores a hard regulatory requirement. Option C misunderstands temperature, which affects randomness in generative models and has nothing to do with explaining a classifier. Option D is false; lending is one of the most heavily regulated domains for explainability.

Explanation

Amazon Macie uses machine learning and pattern matching to discover and classify sensitive data such as PII and financial data in Amazon S3, producing findings that tell you which objects contain what; the exam guide lists it among services that secure AI systems, and preventing sensitive data from leaking into a knowledge base is a data-leakage-prevention concern the 2026 guide added. Option A scans compute workloads and container images for software vulnerabilities, not data content. Option C records API activity. Option D provides AWS compliance reports for download. Recording where the data came from, its lineage, belongs with AWS Glue Data Catalog and Lake Formation.

Explanation

Context engineering, added to the exam guide in 2026, is the discipline of managing everything the model sees in a request: which instructions are current, which retrieved passages are relevant, how much history to keep, and what tool outputs or memory to include. The scenario is a context problem, not a model problem, so curating the context fixes it. Option A bakes documents into weights, which is slow and costly and does nothing about a cluttered prompt. Option C changes randomness, not relevance. Option D may not help at all if the relevant document is still being pushed out; a bigger model reading the wrong context is still wrong.

Explanation

The request requires actions, not just an answer: querying an API, making a booking, and writing to a calendar in the right order. That is the role of an agent, which uses the foundation model to plan a sequence of steps and call tools, then reasons over the results; the exam guide asks you to define agents' role in multi-step tasks and their business applications. Option A would let the assistant answer questions about policy but cannot book anything. Option C might improve how the model talks about bookings but gives it no ability to act. Option D extracts entities from text, a small part of the job, and cannot orchestrate the rest.

Explanation

Regression predicts a continuous numeric value, and kilowatt-hours is continuous, so the consumption forecast is regression. Grouping items by similarity with no labels provided is clustering, an unsupervised technique, so the ticket project is clustering. Option B is wrong because classification assigns discrete categories, not numeric quantities. Option C is wrong because complaint-or-not is a binary classification problem with a defined label, not clustering. Option E is wrong because reinforcement learning optimises actions against a reward signal and has no role in discovering groups in static ticket data.

Explanation

The Model Context Protocol is an open standard that lets an agent discover and invoke tools and data sources through a common interface, so each tool is exposed once as an MCP server rather than integrated by hand into every agent; the 2026 exam guide names it explicitly. Option A, RAG, retrieves documents to ground answers and does not give an agent the ability to take actions in other systems. Option C would teach the model about the APIs but gives it no mechanism to call them, and it would go stale when the APIs change. Option D is a storage pattern for retrieval, not an integration standard for tool use.

Explanation

Prompt caching stores the processed form of a repeated prompt prefix so subsequent requests that begin with the same content skip reprocessing it, which cuts both the input-token charge for the cached portion and the latency before the first output token. The 2026 exam guide lists prompt caching among foundation model selection criteria. Option A changes randomness and neither cost nor speed. Option C buys dedicated capacity and can stabilise throughput, but every request still processes the full 3,000-token preamble. Option D would take a training run to remove a problem a caching feature solves, and the policy would go stale when it changed.

Explanation

SageMaker Model Cards are structured records of a model's purpose, training data, evaluation metrics, risk rating, and limitations, kept alongside the model for governance and audit; the exam guide names them under transparency and explainability tools and again under documenting data origins. AWS publishes the equivalent AI Service Cards for its own managed services. Option A filters content at run time and documents nothing. Option C stores documents for retrieval, not model metadata. Option D records API calls for auditing account activity; it can show who deployed a model but not what the model is for or how it was evaluated.

Explanation

Amazon Transcribe converts speech to text, which produces the subtitles. Amazon Translate performs neural machine translation into the target languages. Amazon Rekognition analyses images and video and includes content moderation that detects unsafe or explicit material. All three are fully managed APIs that need no model training. Option B swaps in Comprehend, which analyses text rather than audio, and Textract, which extracts text from documents rather than moderating images. Option C uses Polly, which is text-to-speech, the reverse of what subtitling needs. Option D adds a custom SageMaker AI model for moderation when a managed capability already exists, and Comprehend cannot translate.

Explanation

The 2026 exam guide added Amazon Bedrock AgentCore Identity and Policy in AgentCore to the list of features that secure AI systems. AgentCore Identity manages the agent's own identity and delegated access to third-party or AWS resources on behalf of a user, so the agent acts with that user's permissions rather than a shared secret. Policy in AgentCore evaluates rules against every tool call that passes through AgentCore Gateway, enforcing constraints such as read-only. Least-privilege IAM behind the tools completes the layering. Option A violates least privilege and makes every injection catastrophic. Option C filters content, not authorisation. Option D is a data-discovery service with no authorisation role.

Explanation

Providing a handful of worked examples of the desired input-to-output mapping inside the prompt is few-shot prompting; one example is single-shot. It steers format and style without any training. Option A, zero-shot, gives an instruction with no examples, which is what the developer had before. Option C, chain-of-thought, asks the model to reason step by step and is used to improve accuracy on multi-step problems, not to fix formatting. Option D, a negative prompt, tells the model what to avoid and is most associated with image generation. The exam guide lists all four under prompt engineering techniques and constructs.

Explanation

The exam guide's agentic AI concepts include memory management, tool usage, multi-agent patterns, and workflow orchestration. Distinguishing session-scoped short-term memory from persistent long-term memory is correct, and an orchestrator or supervisor agent that decomposes a goal and delegates to specialists is the standard multi-agent pattern. Option C is the opposite of what defines an agent: taking actions through tools is central. Option D is false because tools are described to the model at run time, through function definitions or MCP, without changing its weights. Option E is wrong because the foundation model is what decides which tool to call and interprets the result.

Explanation

The 2026 exam guide explicitly tests when a traditional ML model is preferable to a foundation model, and this scenario hits the three named reasons: regulatory concern, an explainability requirement, and structured tabular data. A conventional classifier trained in SageMaker AI is well suited to tabular inputs, and SageMaker Clarify produces feature attributions (SHAP values) that show which inputs drove each decision. Option A gives no reliable feature-level explanation and risks hallucinated reasoning. Option C inherits the same opacity at higher cost. Option D adds orchestration complexity without solving the core problem: an LLM's decision is still not attributable to input features in the way regulators require.

Explanation

Chain-of-thought prompting instructs the model to lay out intermediate reasoning before its conclusion. For problems with several dependent steps this measurably improves accuracy, because the model conditions each step on the previous written one instead of jumping to an answer. Option A does the opposite of what is needed, cutting off the space the model would use to reason. Option C is meaningless; a negative prompt excludes content or themes, and telling a model not to be wrong does not make it right. Option D would produce a smaller, cheaper model, not a more accurate one, and the team said it cannot retrain.

Explanation

A hallucination is a fluent, confident statement that is false or unsupported. The exam guide lists it as a core GenAI disadvantage and, since 2026, asks about detection and grounding methods. Retrieval Augmented Generation supplies the model with verified text to answer from, guardrails can block or flag unsupported claims (Amazon Bedrock Guardrails includes a contextual grounding check), and human review catches the rest. Option A misdiagnoses a generation problem as a training-fit problem. Option C would increase randomness and therefore the risk of invented facts. Option D confuses a malicious-input attack with the model inventing content on its own.

Explanation

The exam guide lists user-feedback mechanisms and AI decision transparency as principles of human-centred design for explainable AI. Showing sources and confidence lets the person judge the answer instead of trusting it blindly, and a feedback loop surfaces failures so the system improves. Option C removes transparency and misleads users, the opposite of responsible practice. Option D takes the human out of a high-stakes loop where human oversight is essential. Option E maximises randomness in a setting that demands consistency and accuracy.

Explanation

The 2026 exam guide added audit-trail and logging requirements for AI interactions. AWS CloudTrail records every Bedrock API call with the caller identity, time, and parameters, which answers who invoked what and when. Model invocation logging, enabled in the Bedrock account settings, writes the full request and response bodies to an S3 bucket or CloudWatch Logs so the content itself can be reviewed and retained. Option A offers best-practice recommendations, not an audit log. Option C supplies AWS's own compliance reports, not records of your activity. Option D evaluates resource configuration and can flag that logging is disabled, but it does not capture the invocations.

Explanation

A negative prompt lists what the output should not contain, and image generation models use it to steer away from unwanted elements such as text, logos, or extra limbs. The exam guide names negative prompts among prompt engineering constructs alongside context and instruction. Option A adds randomness, which makes unwanted artefacts more likely, not less. Option C narrows token sampling in text models and is not the control for excluding visual elements. Option D is a heavy, costly intervention for a problem a single prompt field solves, and the team would still need to describe what to avoid.

Explanation

Stable training metrics with falling production accuracy means the real-world data has moved away from what the model learned, which is drift. The MLOps answer is to monitor production inputs and predictions continuously, for example with SageMaker Model Monitor, and to trigger retraining on fresh data when drift crosses a threshold, rather than on a fixed calendar. Option B changes how hard the model fits the old data and does nothing about the data being stale. Option C changes how predictions are delivered, not how accurate they are. Option D addresses latency or throughput, neither of which is the reported problem.

Explanation

Nondeterminism is a listed limitation of generative AI: the model samples from a probability distribution over tokens, so identical prompts can yield different wording. It is expected behaviour, not a defect. Where consistency matters, reduce randomness by setting temperature near zero and tightening top-p or top-k, and consider constraining output format. Option A misrepresents normal behaviour as a bug. Option C misuses the term hallucination, which refers to false content, not to variation in wording of correct content, and fine-tuning does not remove sampling randomness. Option D affects only answer length and has no bearing on repeatability.

Explanation

The message attempts to override the assistant's instructions and exfiltrate its configuration and data, which is prompt injection; the exam guide lists exposure, poisoning, hijacking, and jailbreaking as prompt engineering risks. Defences layer: Guardrails filter malicious or off-topic input and sensitive output, the application keeps user text clearly separated from system instructions and validates it, and the agent's tools are scoped so that even a successful injection cannot reach records it should not. Option A concerns invented facts, not adversarial input. Option C is a training problem unrelated to runtime attacks. Option D describes changing data patterns, not a deliberate manipulation attempt.

Explanation

Aggregate accuracy can hide harm to a subgroup, which is why the exam guide lists subgroup analysis, human audits, and label-quality analysis as tools to detect and monitor bias, trustworthiness, and truthfulness. Monitoring predictions by region, demographic, or other cohort in production, and scheduling human review of the results, catches the divergence early. SageMaker Clarify computes bias metrics and Model Monitor can track them over time. Option A changes throughput, not fairness. Option C is the mistake that allowed the problem to grow. Option D is irrelevant; text-to-speech has no role in bias detection.

Explanation

The exam guide lists model selection factors such as cost, latency, modality, multilingual support, model size and complexity, customisation options, and input and output length. The stated requirements map directly to two of them: language coverage, and latency and cost at volume. Option C is irrelevant on its own; a newer model is not automatically better for these languages or cheaper. Option D concerns a modality the use case does not need. Option E describes a deployment preference that contradicts the scenario, which is about picking a hosted model for a chat service, and it would add operational cost rather than reduce it.

Explanation

The exam guide lists AWS Config, Amazon Inspector, AWS Artifact, AWS CloudTrail, and AWS Trusted Advisor as services that assist with governance and regulatory compliance. AWS Artifact is the self-service portal for AWS's own audit reports and agreements, which proves the provider side. AWS Config records and evaluates the configuration of your resources against rules, such as requiring encryption on S3 buckets that feed a knowledge base, which proves your side. Options C, D, and E are AI application services for speech, translation, and recommendations, and none of them plays a governance role.

Explanation

This is the managed-API versus self-hosted decision the exam guide names under "methods to use a model in production". Amazon Bedrock exposes foundation models through a serverless API with on-demand token-based pricing, so there is no infrastructure to size, but you take the model as offered. SageMaker JumpStart deploys a model onto an endpoint you own, which gives control over instance type, scaling, and customisation, at the cost of managing that capacity and paying for it whether or not it is busy. Option B reverses the two. Option C is false: Bedrock is serverless. Option D ignores the real differences in control, operational effort, and pricing model.

Explanation

Amazon Bedrock Prompt Management is a central store for prompts with versioning, variables, testing against models, and reuse across applications, which is exactly the governance the scattered templates lack; the 2026 exam guide added prompt versioning and management strategies as an objective. Option A evaluates model outputs against metrics or human judgement and does not manage prompt text. Option C filters harmful or off-topic content in inputs and outputs. Option D provides retrieval over documents for RAG. None of the three tracks prompt versions or gives teams a single source of truth for them.

Explanation

Business value is measured by outcomes the organisation cares about. The exam guide names conversion rate, return on investment, efficiency, average revenue per user, and customer lifetime value as GenAI business metrics, and conversion rate is the natural one for outreach emails. Option A is a usage statistic that says nothing about whether the emails work. Option C, BLEU, measures how closely generated text matches a reference; a persuasive email need not resemble any reference, so the score is a poor proxy for value. Option D is a technical property of the model that leadership has no reason to care about.

Explanation

Recall measures the share of actual positive cases the model catches. When a missed positive is the costly error, as in medical screening or fraud detection, recall is the metric to maximise, accepting more false alarms in exchange. Precision, option A, measures how many flagged cases were truly positive; optimising it would reduce false alarms at the expense of missing real cases, the opposite of what the hospital wants. Accuracy, option C, is misleading when the condition is rare, since a model that flags nobody can score very high. RMSE, option D, is a regression metric for continuous predictions and does not apply to a yes-or-no classification.

Explanation

RLHF collects human judgements comparing candidate responses, trains a reward model on those preferences, and then optimises the language model against that reward, so the model learns qualities such as helpfulness and tone that are hard to capture with a single reference answer. The exam guide lists RLHF under preparing data to fine-tune a foundation model. Option A finds groups in data and trains nothing. Option C teaches domain language from raw text and encodes no preference signal. Option D changes sampling randomness at inference time and cannot change what the model has learned to prefer.

Explanation

The exam guide includes environmental considerations and sustainability among responsible practices for selecting a model. When a smaller model satisfies the requirement, using a much larger one spends energy, money, and latency for no benefit, so the smaller model is the responsible choice; it is also often the more explainable and faster one. Option A treats capability as an end in itself rather than matching the model to the task. Option C is nonsensical; temperature affects sampling randomness, not compute or energy consumption. Option D optimises a property the task does not need.

Explanation

The Generative AI Security Scoping Matrix classifies generative AI use into five scopes, from consuming a public application, through using enterprise apps and pre-trained models, to fine-tuned and self-trained models, and maps the security, governance, compliance, legal, and privacy considerations that grow with each scope. The 2026 exam guide names it as an example governance framework. Option A recommends account best practices and is not a scoping framework. Option C finds software vulnerabilities in workloads. Option D provides compliance reports. None of these helps decide which controls apply to which kind of generative AI project.

Explanation

Amazon Bedrock AgentCore is the managed platform for deploying and operating AI agents at scale. Its components include Runtime for serverless execution, Memory for short- and long-term context, Identity for agent authentication and delegated access, Gateway to expose APIs and Lambda functions as tools, Policy for authorisation, and Observability; it was added to the exam guide in April 2026. Option B is a no-code tool for building traditional ML models. Option C is the business-intelligence service formerly called QuickSight. Option D is a natural-language-processing API for sentiment, entities, and PII detection, with no agent capabilities.

Explanation

ROUGE (Recall-Oriented Understudy for Gisting Evaluation) measures how much of a reference summary's content the generated summary recovers and is the conventional metric for summarisation. BLEU (Bilingual Evaluation Understudy) measures n-gram precision against reference translations and is the conventional metric for machine translation. BERTScore, also in the exam guide, compares semantic similarity using embeddings and works for either. Option B swaps the two. Option C uses a regression error metric that does not apply to text generation. Option D treats open-ended text as if it had a single correct label, which it does not.

Explanation

The exam guide separates model performance metrics from business metrics and expects you to tell them apart. Cost per resolved ticket and customer satisfaction measure the effect of the assistant on the business: money spent per outcome and how customers feel about the result. F1 score and precision are classification quality metrics computed against labelled data, and ROUGE measures how closely generated text overlaps a reference summary; all three describe the model, not the business outcome. Other business metrics the guide names include return on investment, conversion rate, average revenue per user, and customer lifetime value.

Explanation

Strands Agents is AWS's open-source, model-driven SDK: developers describe an agent's model, tools, and prompt in a few lines of code, and the framework handles the reasoning loop. Amazon Bedrock AgentCore is the managed infrastructure layer where such agents, built with Strands or other frameworks, run in production with runtime, memory, identity, gateway, and observability handled by AWS. Both appear in the 2026 exam guide as GenAI development offerings. Option B reverses the roles. Option C is wrong because neither is a model; Amazon's own models are the Nova family. Option D invents a modality split that does not exist.

Explanation

LLM-as-a-judge uses a capable foundation model to grade outputs against defined criteria, which scales to large volumes and handles open-ended answers where no reference exists; the 2026 exam guide added it alongside ROUGE, BLEU, and BERTScore. Amazon Bedrock Model Evaluation supports automatic, human, and judge-model evaluations. A human-reviewed sample checks that the judge agrees with people. Option A needs a reference answer per question, which the scenario rules out, and BLEU measures overlap rather than helpfulness. Option C measures length, not quality. Option D is a regression metric with no meaning for text.

Explanation

The exam guide names inclusivity, diversity, curated sources, and balance as characteristics of responsible datasets, and describes how bias in data affects demographic groups. A convenience sample under-represents some speakers, so the model sees too few examples of their speech and performs badly for them. The fix is deliberate curation: collect or source recordings covering the missing group so the data represents the users the model will serve. Option A would worsen coverage. Option C throws away the metadata needed for subgroup analysis. Option D is irrelevant to a recognition model and does nothing about data coverage.

Explanation

The 2026 exam guide added hallucination detection methods and grounding techniques: RAG grounding, output validation, and confidence scoring. A contextual grounding check compares the generated answer with the retrieved passages and flags or blocks claims the sources do not support; Amazon Bedrock Guardrails provides this as a configurable filter. Pairing it with confidence thresholds and human review for borderline cases completes the control. Option A increases the chance of invented content. Option C removes the very source material that makes grounding possible. Option D adds tokens without adding any verification step.

Explanation

Kiro is AWS's agentic IDE: an AI coding environment that plans work from specifications and writes and refactors code, and it entered the exam guide in the April 2026 revision. Amazon Quick, the renamed Amazon QuickSight, lets business users ask questions in natural language and get answers, dashboards, and visualisations. Option B misapplies AWS Transform, which is an agentic service for migrating and modernising legacy workloads such as mainframe or .NET applications. Option C pairs SageMaker Canvas, a no-code tool for building ML models, with Knowledge Bases, a RAG building block, neither of which is an IDE assistant or a BI tool. Option D names text-analysis and recommendation services that fit neither need.

Explanation

Provisioned Throughput buys dedicated model capacity, measured in model units, for a term commitment, which gives guaranteed throughput and consistent performance for steady, high-volume traffic and avoids throttling. On-demand pricing charges per input and output token with no commitment, which suits sporadic or unpredictable usage where reserved capacity would sit idle. Option B pays for dedicated capacity the internal tool would rarely use while leaving the busy application exposed to throttling. Option C wastes money on the low-volume tool. Option D describes a pricing model Bedrock does not offer; foundation model access is metered, not licensed outright.

Explanation

The exam guide distinguishes evaluating a model from evaluating an application built with it, and lists task completion rate, user satisfaction, and cost per interaction as business-objective alignment metrics. Completion rate shows whether the assistant does the job, and cost per interaction shows whether it does so economically. Options C, D, and E are properties of the model and embedding pipeline that say nothing about outcomes: a larger model, a longer embedding vector, or a lower training loss can all coexist with an assistant that fails to resolve tickets or costs more than the people it replaced.