CCAR-F Practice Questions: Context Management & Reliability Domain

Correct answer (C): Explicit triggers create consistent escalation behavior, and a structured handoff preserves the reason, known facts, attempted actions, and remaining need. The underlying principle is that escalation must be policy-driven and context-preserving rather than based on vague judgment. In production, this improves routing consistency and prevents customers or reviewers from reconstructing the case.

Why the other options are wrong:
- Option A: Negative sentiment alone does not establish the need for escalation, and a complete transcript may burden the reviewer with unnecessary detail.
- Option B: Implicit inference from tone produces inconsistent triggers, while the last message alone provides an incomplete handoff.
- Option D: An arbitrary turn limit is not tied to a human request, policy gap, or actual inability to progress.

Correct answer (C): The structured payload gives the coordinator the failure type, attempted action, useful partial results, and recovery options. The CCAR-F principle is to propagate actionable error state rather than collapse failures into generic status values or discard completed work. In production, this enables targeted retries, alternative sourcing, and accurate disclosure of remaining coverage gaps.

Why the other options are wrong:
- Option A: A generic flag is easy to process, but it does not distinguish a recoverable rate limit from a permanent or unknown failure.
- Option B: Preserving partial results is valuable, but omitting the failure prevents the coordinator from understanding why coverage is incomplete.
- Option D: Raw logs can support deep debugging, but they create unnecessary context load and are less actionable than a concise structured failure contract.

Correct answer (D): A durable manifest provides examined components, findings, unresolved questions, and next actions, while targeted validation accounts for repository changes since those findings were recorded. The CCAR-F principle is recoverable exploration based on structured checkpoints rather than raw session history. In production, this avoids both blind trust in stale state and the cost of repeating an entire investigation.

Why the other options are wrong:
- Option A: A transcript may contain useful evidence, but it is noisy, and following every prior conclusion ignores changes that may invalidate findings.
- Option B: A full scan discards valid checkpointed work and fails the requirement to resume efficiently.
- Option C: A patch does not reliably capture rejected paths, unresolved questions, or the architectural evidence gathered.

Correct answer (C): Stratified evaluation exposes performance on rare document types and sensitive fields, while calibrated field-level thresholds support targeted review. The underlying principle is that aggregate accuracy and document-level confidence can hide consequential subgroup and field failures. In production, risk-aware calibration directs human effort where errors matter most and allows thresholds to adapt to measured performance.

Why the other options are wrong:
- Option A: Rare document types may be underrepresented in a simple random sample, and one overall score hides field-specific risk.
- Option B: A document-level threshold may overlook an overconfident error in a high-risk field.
- Option D: This addresses one known segment but ignores calibration drift and other document types or fields that may also be unreliable.

Correct answer (D): Both figures should retain their source attribution and dates, with the synthesis explaining that they may describe different points in time rather than a direct factual contradiction. The underlying principle is to preserve conflicting or changing evidence instead of silently collapsing it based on recency. In production, this maintains provenance, accurately represents uncertainty, and prevents valid historical data from being erased.

Why the other options are wrong:
- Option A: The newer source may best describe the latest state, but it does not invalidate the older figure in a report about trends over time.
- Option B: These dated figures likely represent different workforce states and should not be averaged.
- Option C: Omitting dates removes the information most useful for understanding the apparent disagreement.

Correct answer (B): A structured summary preserves decision-critical identifiers, amounts, dates, and unresolved issues while removing repetitive content. The CCAR-F principle is to compact context without discarding exact persistent facts. In production, this reduces context pressure while preventing incorrect actions caused by vague or missing case details.

Why the other options are wrong:
- Option A: Keeping the entire raw history increases context size and can make important facts less salient.
- Option C: The narrative removes the exact amount, date, identifier, and investigation status needed for reliable handling.
- Option D: Discarding earlier messages can remove binding case details that remain relevant even if they are not recent.

Correct answer (B): A structured case summary preserves exact amounts, identifiers, dates, and unresolved issues while allowing verbose tool output to be removed. The CCAR-F principle is to compact context around durable, decision-relevant state rather than preserve an undifferentiated transcript. In production, this reduces token consumption and lost-in-the-middle risk without sacrificing facts needed for correct case handling.

Why the other options are wrong:
- Option A: Keeping raw responses can preserve evidence, but retaining all verbose output wastes context, and removing customer messages may discard relevant intent or commitments.
- Option C: A narrative summary is compact, but omitting exact identifiers and values can cause lookup errors or unsupported decisions.
- Option D: Reordering one response may improve its prominence, but it does not control context growth or explicitly preserve all critical facts.

Correct answer (B): Field-level calibration and stratified sampling can expose weak customs-code extraction and performance gaps in rare handwritten invoices. The underlying principle is that aggregate metrics and document-level confidence can conceal poorly performing segments. In production, segment-aware review directs scarce human effort toward risks that broad averages would miss.

Why the other options are wrong:
- Option A: Low-confidence routing is useful, but one document-level score can hide a high-risk field and may be poorly calibrated for rare document types.
- Option C: Simple random sampling estimates broad performance, but rare invoice types may be underrepresented and field-specific failures can remain hidden.
- Option D: A deployment threshold can be part of governance, but aggregate accuracy alone does not establish safety for every field and segment.

Correct answer (B): Distinct states preserve the semantic difference between finding no records and being unable to determine whether records exist. The CCAR-F principle is that error contracts must not conflate absence of evidence with evidence of absence. In production, explicit outcomes prevent unsupported customer claims and enable safe fallback retrieval or escalation.

Why the other options are wrong:
- Option A: A shared empty array simplifies consumers, but it falsely represents an access failure as a completed search with no matches.
- Option C: Retries can recover transient errors, but repeatedly retrying a permission failure is unlikely to help and does not define a truthful result contract.
- Option D: An unknown value acknowledges uncertainty, but it omits the failure cause and forces downstream inference instead of providing actionable state.

Correct answer (B): The subagent can safely attempt the documented alternative locally and, if recovery fails, provide the coordinator with the failures, attempted actions, and partial results. The underlying principle is to recover locally when a safe path exists while preserving structured error information for higher-level decisions. In production, this avoids unnecessary escalation but prevents failed access from being misrepresented as proof that an order does not exist.

Why the other options are wrong:
- Option A: Immediate propagation preserves the original error but ignores an explicitly safe local recovery path.
- Option C: A validation error requires changing the request, and returning an empty object would conceal the failure.
- Option D: Conversation history is not authoritative enough to replace a required system-of-record lookup.