CCAR-P facts to memorise: 30 recall cards
Thirty facts that are pure recall — the ones you either know cold or lose a mark on. Cover the right-hand column and work down the list.
| Prompt | What you need to know |
|---|---|
| Exam format | 63 scored questions, 120 minutes, 720/1,000 scaled. Standalone items, not scenario-anchored. No prerequisite. |
| Heaviest domain | Integration, 19% — ahead of Solution Design at 17%. |
| The new 35% | Governance and Risk (14%), Stakeholder and Lifecycle (14%), Developer Enablement (7%) have no Foundations equivalent. |
| The defensibility test | Every decision needs a stated trade-off, a measurement, and an owner. |
| Before choosing a pattern | Establish stakes, hard constraints (residency, latency, retention, regulation) and the success measure. |
| Workflow vs agent, at this level | Agentic systems are harder to bound, audit and cost-forecast — often decisive in a regulated context. |
| Guardrail layers, strongest first | Architectural · deterministic validation · model screening · prompt-level · human review. |
| Identity in integrations | Propagate the caller's identity to the data layer. One privileged service account means any user can get any row. |
| Deployment surface choice | A compliance decision: IAM-based auth, existing billing, invocation logging, and regional endpoints for residency. |
| RAG chunking failure | Chunks stripped of context — "revenue grew 3%" with no company and no quarter. |
| Contextual enrichment | Prepend a short generated description situating each chunk in its document. Prompt caching makes it economic. |
| Why hybrid retrieval | Embeddings miss exact identifiers and rare terms; BM25 misses paraphrase. |
| Reranking | Retrieve broadly, score, keep the best. Raises relevance, adds latency, can cut generation cost. |
| Retrieval improvement figures | Contextual enrichment + hybrid ≈ 49% fewer top-20 retrieval failures; adding reranking ≈ 67%. |
| Citations | Structured pointers: cited_text, document_index, char/page/block location. cited_text is free on output. Not compatible with strict structured output. |
| Hallucination in RAG | Check retrieval first — often the model answered faithfully from chunks lacking the answer. |
| Observability per request | request_id · model and prompt version · stop_reason · full usage · TTFT and total · tool calls · retrieved vs cited · validation and guardrail outcomes · escalations · cost attribution. |
| Earliest drift signal | Human override and escalation rate — infrastructure metrics stay flat while quality falls. |
| Safe change sequence | Offline evals → shadow run → A/B with limited exposure → rollout. One variable at a time. |
| Shadow vs A/B | Shadow gives production distribution with zero user exposure; A/B gives real outcomes at the cost of exposure. |
| LLM-as-judge | Validate against human labels first. Specific empirical rubric, parseable output, reasoning before score. |
| Eval volume | 100+ cases; 1,000+ for production. Refresh from real traffic as the distribution shifts. |
| Anthropic certifications | SOC 2 Type I and II · ISO/IEC 27001:2022 · ISO/IEC 42001:2023 (AI management) · HIPAA-ready with BAA. |
| Vendor certification | An input to your compliance case, never a substitute. Lawful basis, minimisation, retention, access control and audit evidence stay yours. |
| Commercial data handling | Team, Enterprise and API content not used for training by default; zero data retention arrangements available. |
| High-risk domains | Legal · healthcare · insurance · financial advice · employment · housing and lending · academic assessment · journalism. |
| Human review design | Triage which cases · give reviewers evidence and citations · measure override rate · size the queue. A backlogged review is no review. |
| Over-refusal | A quality failure, not a safety win — users route around a system that declines legitimate work. |
| SLA: can commit | Availability · latency percentiles · accuracy against a defined eval set · escalation turnaround · time to detect degradation. |
| SLA: cannot commit | Correctness on arbitrary inputs · that degradation will not occur. Your availability is bounded by provider + retrieval + your own infrastructure. |
← Previous: Questions & answers — Domains 6–7 | Next: Six-week study plan →
Explore Other Certification Study Guides
Switch to another exam guide to prepare for your next Claude certification:
Associate
CCAO-F
Claude Certified Associate — Foundations
Entry-level certification for non-coders: prompting, output evaluation, model selection, workflow integration, configuration, governance & responsible use.
Open Guide →
7 Domains
Architect
CCAR-F
Claude Certified Architect — Foundations
Core architectural principles: agentic loops & orchestration, tool design & MCP, Claude Code workflows, prompt engineering, and context reliability.
Open Guide →
5 Domains
Developer
CCDV-F
Claude Certified Developer — Foundations
Build software on Claude: API integration, model selection, tools & MCP servers, agents vs workflows, security, safety & debugging.
Open Guide →
8 Domains