CCAR-P cheat sheet
What to hold in memory the morning of the exam.
Exam facts
| Scored questions | 63 | Time | 120 minutes (~1.9 min each) |
|---|---|---|---|
| Pass mark | 720 / 1,000 scaled | Format | Standalone items, multiple choice and multiple response |
| Prerequisites | None | Validity | 12 months |
Domain weights, in order
| Weight | Domain |
|---|---|
| 19% | Integration |
| 17% | Solution Design and Architecture |
| 16% | Evaluation, Testing and Optimisation |
| 14% | Governance, Safety and Risk |
| 14% | Stakeholder Communication and Lifecycle |
| 13% | Claude Models, Prompting and Context Engineering |
| 7% | Developer Productivity and Operational Enablement |
The three domains with no Foundations equivalent — governance, stakeholder and enablement — total 35%.
The defensibility test
Every design decision needs a stated trade-off, a measurement, and an owner. An option missing any of the three is usually the distractor.
Guardrail layers, strongest first
- Architectural — the capability does not exist
- Deterministic validation — schemas, business rules, policy in code
- Model-based screening — cheap classifier over inputs and tool outputs
- Prompt-level — hardened prompts, untrusted-content policy
- Human review — for the cases the stakes demand
RAG pipeline and its failure at each stage
| Stage | Fails as |
|---|---|
| Chunking | Context-stripped chunks |
| Contextual enrichment | Skipped — chunking failure persists |
| Embedding search | Misses exact identifiers and rare terms |
| Lexical (BM25) | Misses paraphrase — hence hybrid |
| Reranking | Omitted — marginal chunks distract and inflate cost |
| Citations | Absent — claims unverifiable |
Published figures: contextual enrichment plus hybrid retrieval cuts top-20 retrieval failure by ~49%; adding reranking takes it to ~67%.
Citations
Enable per document; returns cited_text, document_index and a character, page
or block location. More reliable than prompting for citations, and cited_text does not count
toward output tokens. Not compatible with strict structured output.
Compliance quick reference
| Anthropic holds | SOC 2 Type I and II · ISO/IEC 27001:2022 · ISO/IEC 42001:2023 (AI management) · HIPAA-ready with a BAA available |
|---|---|
| Commercial data | Team, Enterprise and API content is not used to train models by default; zero data retention arrangements available |
| Data residency | Cloud deployment surfaces offer regional endpoints, typically at a premium |
| Still yours | Lawful basis, minimisation, retention, access control, audit evidence |
High-risk domains needing human review and disclosure
Legal · Healthcare · Insurance · Financial advice · Employment · Housing and lending · Academic assessment · Journalism.
Observability: log per request
request_id · model and version · prompt version · stop_reason ·
full usage including cache fields · TTFT and total latency · tool calls and errors ·
what was retrieved and what was cited · validation and guardrail outcomes · escalations ·
cost attributed to tenant or feature.
Safe change sequence
Offline evals → shadow run on real traffic → A/B with limited exposure → full rollout. One variable at a time. Watch cost and latency alongside quality.
What you can and cannot put in an SLA
| Can | Availability, latency percentiles, accuracy against a defined eval set, escalation turnaround, time to detect degradation |
|---|---|
| Cannot | Correctness guarantees on arbitrary inputs, or that degradation will not occur |
Carried over from Foundations, still examinable
Branch on stop_reason, never on prose · schemas constrain shape, not truth ·
4 cache breakpoints ordered tools → system → messages,
reads 0.1× and exempt from ITPM · Batches 50% off but asynchronous up to 24 h ·
retry 429/500/504/529, never 400/401/403/404/413 · subagents inherit only the prompt string ·
CLAUDE.md guides, settings and hooks enforce.
← Previous: The 14 anti-patterns | Next: Questions & answers — Domain 1 →
Explore Other Certification Study Guides
Switch to another exam guide to prepare for your next Claude certification:
Claude Certified Associate — Foundations
Entry-level certification for non-coders: prompting, output evaluation, model selection, workflow integration, configuration, governance & responsible use.
Claude Certified Architect — Foundations
Core architectural principles: agentic loops & orchestration, tool design & MCP, Claude Code workflows, prompt engineering, and context reliability.
Claude Certified Developer — Foundations
Build software on Claude: API integration, model selection, tools & MCP servers, agents vs workflows, security, safety & debugging.