CCA Study Tips for Domain 5: Context Management & Reliability

Study Domain 5 (Context Management & Reliability, 15% of the CCA-F) last, in four sessions, by reading the 31 articles in task order and grouping them under six one-line principles you can recite: numbers die in summaries, attention is U-shaped, escalate on observable conditions, errors need structure between components, specificity decays with time, and aggregates hide failures. This post gives the reading order, the memorise-versus-reason split, a practice cadence, self-checks, and the mistakes to review twice.

For what the domain covers and why it is weighted at 15%, read CCA Domain 5 Explained. This post is only about how to study it.

What order should I read the Domain 5 articles in?

Task order on the Domain 5 study guide page. Suggested sessions:

Session	Articles	Goal
1	Task 5.1: “$127.50 Refund” Became “Customer Requested a Refund”, The U-Shaped Attention Curve, 40 Fields Returned, 5 Needed, The API Does Not Remember Your Conversation, the case-facts tracker, 28% of Issues Dropped Without a Tracker, 2,800 Tokens of Reasoning Chain → 280 Tokens of Structured Facts	Write a case_facts block for a sample support conversation
2	Tasks 5.2 and 5.3: Sentiment-Based Escalation: Replace It, immediate escalation, The Signal Is Broken, multi-match disambiguation, policy gaps, Generic “Failed” → 18%. Structured Error → 71%., the decommissioned database, silent swallow, access failure vs empty, coverage annotation	List the four reliable triggers and the two unreliable ones without notes
3	Task 5.4: Minute 0 vs Minute 45, The Scratchpad, Let Someone Else Hold the Data, Crash Recovery Manifest, Save, Compact, Restore	Describe the four countermeasures and when each applies
4	Tasks 5.5 and 5.6: Why Aggregate Metrics Lie, Stratified Sampling, Field-Level Confidence, calibration, Every Fact Needs a Return Address, Present Both, Fabricate Neither, temporal metadata, established vs contested, format matching	Given two conflicting figures, write the correct one-sentence output

If you have not studied Domains 1 and 2 yet, do them first. Domain 5 questions assume you know sub-agent context isolation and the isError distinction.

What should I memorise, and what should I reason about?

Memorise:

• Summarization retention direction: sentiment and categories survive; exact amounts, dates and IDs mostly do not. Fix: a persistent structured facts block that is never compressed.
• Attention is U-shaped; middle sources are incorporated about half as often; window size does not fix it; headers plus a leading summary do.
• Reliable escalation triggers: explicit human request, policy gap, no progress after two attempts, policy exception needed. Unreliable: negative sentiment, low model confidence.
• Two matching customer records means ask for a disambiguator, never auto-select (Auto-Select “Most Recent”: 27% Wrong Customer).
• Access failure is an error, not an empty result, on both sides of the interface.
• The /compact workflow: save findings to a file, compact, restore by reading the file.
• Conflicts are preserved with attribution and dates; averaging is fabrication.

Reason about:

• What did the summary drop? Look for a downstream action that used a wrong number or ID.
• Is this escalation trigger observable? If it depends on emotion or a confidence score, it is not.
• Where should the error be handled? Locally if transient and bounded; propagated with structure if it changes what the orchestrator should do (Two Anti-Patterns That Compound).
• Is this degradation or a hard task? If specificity fell with time and not with difficulty, it is degradation; apply scratchpad and delegation.
• What does the aggregate hide? Disaggregate by category and field before trusting a headline number.
• Is this a contradiction or a difference in time, scope, or method? Check the metadata before calling it a conflict (Not Every Difference Is a Contradiction).

As with the other domains, article-title percentages illustrate direction and are not exam facts.

What practice cadence works for Domain 5?

• After each session, do 15 Domain 5 practice questions. Before revealing, name the principle from the six-item list at the top of this post. If the correct answer relies on a principle you did not name, reread that task.
• After session 4, do a 30-question block. Domain 5 scenarios are longer than average, so use this block to practise reading for the symptom’s timeline (turn 30, minute 45, after summarization) rather than the system’s design.
• Review mistakes weekly. Common ones: escalating on sentiment, believing bigger context fixes attention, and averaging conflicting figures.
• Final week: the free CCA-F mock exam will include roughly 9 Domain 5 questions among 60 (confirm current weights on the official Anthropic / Pearson VUE page before you register).

Self-check questions before moving on

1. After summarization the agent processed a $50 refund instead of $127.50. What was lost and how do you prevent it?
2. Ten sources are concatenated. Which positions are at risk, and what two changes help most?
3. A customer says “I want a person” on their first message. What does the agent do, and why not try to resolve first? (Immediate Escalation → CSAT 3.8)
4. A sub-agent reports “database error”; the orchestrator retries five times against a decommissioned system. What was missing? (The Database Was Permanently Decommissioned)
5. A report says “no studies exist” because a search timed out. What should the report have said instead?
6. By minute 45 an exploration agent contradicts its own earlier findings. Name two countermeasures.
7. Overall accuracy is 97% but one 10% category is at 72%. How should monitoring have been designed?
8. Two sources say 35% and 42%. What is the correct output?

Mistakes worth reviewing twice

• Increasing the context window as a fix for lost-in-the-middle.
• Treating “customer seems angry” as an escalation rule.
• Trusting a 0.9 confidence score without calibration data (The Model Says 0.9 — What Does That Actually Mean?).
• Random-sampling for quality monitoring so rare categories are never checked.
• Choosing the “most recent” or “most authoritative” figure instead of showing both.
• Writing analysis in a table or numbers in prose (Tables for Numbers, Prose for Analysis).

Next step

Once Domains 1 and 2 are done, start session 1 on the Domain 5 study guide page and finish the Domain 5 practice set within the week. Because it is the last domain, follow it directly with the free CCA-F mock exam, then use the domain weighting strategy to decide where the remaining review time goes.