Context Management & Reliability is the smallest domain on the CCA-F at 15%, but it is often the one that surprises candidates. Questions here test your understanding of how Claude handles long contexts, how to design multi-turn sessions that stay coherent across many exchanges, how to calibrate the model's confidence signals, and how the safety model works. These are the topics that separate architects who have shipped production systems from those who have only built demos.
Long-context pitfalls arise when the context window fills up or when important information is buried too deep in a long conversation. Research on large language models consistently shows that model performance degrades on information placed in the middle of a very long context (sometimes called the 'lost in the middle' effect). The exam tests whether you know how to structure context to mitigate this: putting the most important instructions and constraints at the beginning and end of the context, using retrieval to inject only the relevant document chunks rather than entire documents, and summarising earlier turns when a conversation grows long.
Handoff patterns are the techniques you use to preserve session state across context resets. When a conversation exceeds the context window, you cannot simply continue — you need to extract the essential state from the conversation and inject it into a new context. A well-designed handoff prompt captures: the original goal, the decisions made so far, the current state of any external resources that were modified, and any constraints the user established during the session. The exam will present handoff prompts and ask you to identify which one will cause the model to lose track of a previously agreed constraint.
Confidence calibration is about designing your system to surface the model's uncertainty appropriately rather than hiding it. The model will express uncertainty through hedged language ('I believe,' 'you may want to verify'). Applications that strip this hedging before presenting results to users are hiding information that affects user decisions. The exam favours designs that surface uncertainty clearly and that route low-confidence outputs to a human reviewer rather than passing them directly to downstream systems.
The safety model questions test your understanding of Claude's built-in refusal behaviours and how they interact with your application design. Claude will decline certain requests regardless of the system prompt. The exam will present scenarios where an application's expected behaviour conflicts with Claude's safety training and ask you to choose the correct architectural response — which is always to redesign the application task rather than to try to override the safety behaviour. Understanding which constraints are operator-configurable versus hardcoded is a key topic.
Scenario: A legal technology company is building a contract review assistant. The assistant reads a 200-page contract (injected in full into the context), then answers questions about it. In testing, the assistant answers questions about clauses in the first 10 pages and the last 10 pages reliably, but consistently misses or misinterprets clauses in pages 80 to 150.
This is a textbook lost-in-the-middle failure. The correct fix is to replace whole-document injection with a retrieval layer: chunk the contract into sections, embed them, and at query time retrieve the top-k most relevant sections to inject alongside the question. This ensures the relevant clause is always near the current question in the context, regardless of where it appears in the original document. The distractor answers — using a model with a larger context window, lowering the temperature, and splitting the contract into two separate conversations — are all partial workarounds that do not address the structural cause. A larger context window in particular is a very common wrong answer: it delays the problem without solving it.