Integration — study roadmap

This unit is CCARP-U3 · Integration, the largest domain on the exam blueprint at 19%. It carries eight official objectives across four topics. Work them in the order below: each topic assumes the vocabulary of the one before it.

Topic	Objectives	What you must be able to do
T1 · Tool and agent configuration	LO1	Evaluate a configuration for capability bloat and choose the mechanism that targets the measured pressure
T2 · Security analysis	LO2	Separate a working connection from a granted permission, and name the missing control
T3 · Latency and observability	LO3, LO4	Measure the whole path, justify a configuration trade, and instrument a run without over-collecting
T4 · RAG, retrieval and integration mechanisms	LO5–LO8	Design a citable retrieval pipeline, match the matcher to the query, choose the connection mechanism from constraints, and price progressive discovery

How the assets fit together

Topic lectures teach the distinctions and carry a worked example per objective. Watch or read one before its quiz.
Companion quizzes are teaching checks tied to the lecture revision. They are deliberately separate from the exam bank, so answering them does not consume mock-exam questions.
Flashcards hold the distinctions that are easy to blur under time pressure: caching versus context, perceived versus completion latency, retrieval quality versus answer quality.
The exam bank holds 53 questions for this domain, sized from the published 19% weight.
The projects are house-authored exercises. Stage 1 lives with the T1 pilot; the remaining stages are in the domain project.

The four confusions this domain tests

Every one of these appears in the bank more than once, because each is a place where a plausible-sounding claim is wrong.

A connection is not a permission. An accepted authorization token establishes the transport identity of your application. It decides nothing about which tenant, record or operation this call may reach.
Cheaper is not smaller. Prompt caching reduces what you pay for repeated input. The cached tokens still occupy the context window, so a lower bill is not evidence that capability bloat was reduced.
Perceived is not complete. Streaming shows output as it is produced and improves how responsive an application feels. It does not change when the full response finishes.
Retrieved is not grounded. Returning the right chunk and answering only from that chunk are two outcomes, measured with two different instruments.

Sequence

T1 first, because its mechanism table — deferred discovery, caching, context editing, programmatic calling — recurs in T4's progressive-discovery objective.
T2 next. Its layer decomposition is reused in T3, where end-user identity attribution turns out to be an observability problem with a security consequence.
T3 before T4, because T4's retrieval pipeline is something you will be asked to measure, and the measurement vocabulary is defined here.
T4 last and longest. Budget roughly twice the time of any other topic: four objectives, eight house skills, and the only place where chunking, matching, evaluation, protocol choice and discovery are assessed together.

What to carry into the exam

For each objective, be able to state the decision, the evidence you would need, and the claim the evidence does not support. The third of those is what the Professional-level items reward. Several questions in this bank are built from a true measurement paired with an unjustified conclusion drawn from it.

Sources and their limits

Every factual claim in this unit is grounded in retained Anthropic documentation, quoted verbatim with a recorded retrieval date and content hash. Two cautions apply throughout:

Published benchmark figures carry the scope their source gives them, and the scopes differ. The contextual-retrieval cookbook's Pass@k numbers come from its own nine-codebase evaluation set; its headline 35% failure reduction is an average across all the data sources it tested. Both are evidence from that work, neither is a guarantee for another corpus, and they must not be quoted interchangeably.
Beta surfaces change. Trace span names, connector beta headers and platform availability were current at the retrieval date recorded with each snapshot. Re-check them before implementing.

All synthetic numbers in the lectures, cards, questions and projects are house-authored exercise inputs, labelled as such wherever they appear. They exist to make a decision checkable, and they are not vendor measurements.