Exam-style
An internal audit sends one email to a team running a document assistant on Amazon Bedrock and a risk-scoring model trained on Amazon SageMaker AI. It asks four things. Which passage did this specific sentence in the assistant's answer come from? Which ingestion run loaded that passage into the corpus, and what did it do to it on the way in? Which datasets is the corpus assembled from, and who owns each one? And how was the deployed risk-scoring model evaluated, and who approved it? Which four artefacts answer them, in that order?
Reveal the answer
B. A source citation, a data lineage record, a data catalogue entry in the AWS Glue Data Catalog governed with AWS Lake Formation, and an Amazon SageMaker Model Card
Sort the four questions by the size of the thing being asked about, and each one lands on exactly one artefact. The first is about one answer, so it is a source citation: the link an Amazon Bedrock Knowledge Base returns alongside a generated response, naming the retrieved passage the sentence was written from. The second is about one data item’s journey, so it is data lineage: the record of where a piece of data came from and every transformation it passed through, which is what turns “the corpus holds a superseded draft” into a named run on a named date. The third is about the inventory rather than any single item, so it is data cataloging: on AWS, an AWS Glue Data Catalog entry holding the schema and location, with the owner, sensitivity and licence filled in by people and access granted centrally through AWS Lake Formation. The fourth is about the model itself, so it is an Amazon SageMaker Model Card, the versioned document recording intended use, training data, evaluation results, known limitations and the approval that let it run. The swapped option fails because a catalogue entry describes what a dataset is and never what happened to one record inside it, while lineage traces one record and never enumerates the collection. CloudTrail records who called which API and when, which is a useful third leg of an audit and not a link between a sentence and a passage. Folding the first two together fails because a citation reaches back as far as the document and stops: it can say the sentence came from report 0412 and cannot say which version, when it was loaded, or whether a newer one exists. The all-four-on-one-card answer is the one people actually pick, because Model Cards sound the most official of the four, and it is wrong in the same way for each of the first three: a document about a model says nothing about a single answer, a single record, or a catalogue of datasets.
Q. One audit email asks which passage a sentence came from, which run loaded that passage, which datasets the corpus is built from and who owns them, and how the deployed model was evaluated and approved. Which four artefacts?
A. A source citation, a data lineage record, a data catalogue entry, and an Amazon SageMaker Model Card, in that order.
Why? The discriminator is the scope of the thing being described. One answer gets a source citation, produced at answer time by the knowledge base and pointing at the passage the sentence was written from. One data item’s journey gets data lineage, the record of where that item came from and what each step did to it. The inventory of datasets gets the catalogue, which on AWS means data cataloging in the AWS Glue Data Catalog with owners and licence terms filled in and access governed through AWS Lake Formation. The model gets an Amazon SageMaker Model Card, carrying intended use, training data, evaluation results, limitations and the approval. Four scopes, four artefacts, and each one is blind exactly where the next one sees. The trap is reaching for the Model Card every time because it reads as the most formal document in the set, when three of the four questions are not about a model at all.