CCAR-P · Architect Professional — Domain Notes
Cleared Foundations, or aiming straight for the senior badge? CCAR-P trades scenario theater for senior-architect judgment: 63 traditional questions · 120 min · $175 · 720/1000, weighted toward integration, lifecycle, governance, and the consulting dimension — the "in the room with the CTO" material.
Verified August 2026 — re-check the official CCAR-P guide before scheduling.
Blueprint
| Domain | Weight | Core content |
|---|---|---|
| Integration | 19% | deployment platforms, enterprise systems, auth, data flows |
| Solution Design & Architecture | 17% | end-to-end designs, trade-off defense, scalability |
| Evaluation, Testing & Optimization | 16% | eval programs, regression gates, cost/latency tuning |
| Governance, Safety & Risk Management | 14% | Responsible AI (RAI) policy, data handling, auditability, compliance |
| Stakeholder Communication & Lifecycle | 14% | the consulting layer — exec framing, adoption, change management |
| Claude Models, Prompting & Context | 13% | model selection, prompting at depth (the Foundations prompting material, at greater depth) |
| Developer Productivity & Enablement | 7% | rolling Claude Code out to teams, standards, training |
Integration (19%) — the deployment triangle
| Path | When | Architect notes |
|---|---|---|
| Anthropic API (first-party, 1P) | fastest features, direct billing | newest models/features land here first |
| AWS Bedrock | AWS-committed enterprises | Identity and Access Management (IAM) auth, Virtual Private Cloud (VPC) endpoints, AWS-region residency, consolidated AWS billing |
| GCP Vertex AI | GCP-committed enterprises | service-account auth, region pinning, Google billing |
Same models, different auth/billing/residency envelopes; feature lag can exist off-1P — check before promising a capability. Abstract the provider behind your own client interface so switching is config, not rewrite. Enterprise wiring: single sign-on (SSO) / OpenID Connect (OIDC) for humans, scoped service credentials for workloads, secrets in a manager (never code), egress control for data-sensitive deployments.
Solution Design (17%)
The tested skill is trade-off articulation: latency vs cost vs quality (model tiering), build vs buy (MCP ecosystem vs custom tools), workflow vs agent, single vs multi-agent, sync vs batch. Expect "design under constraint" stems where one stated constraint (data residency, cost ceiling, no-human-available) eliminates three plausible answers — the same trap family as Foundations, at system scope.
Evaluation & Optimization (16%)
- Eval program, not eval file: golden sets per capability, LLM-judge rubrics with different-family judges, human calibration samples, regression gates wired into CI, drift monitoring in prod.
- Negative controls (a config mutation must flip a test) prove gates are load-bearing.
- Optimization order: measure → cache → tier models → trim context → batch offline → only then consider fine-tuning-class interventions.
Governance, Safety & Risk (14%)
- Map to frameworks: usage policies, data-retention posture, personally identifiable information (PII) minimization (redact before persistence, not after), audit trails that reconstruct any decision, human accountability for high-impact actions.
- Risk register for agentic systems: injection, tool misuse, data exfiltration via chained tools, runaway cost, hallucinated commitments — each with a deterministic mitigation and an owner.
- Anthropic's RAI posture (safety filtering, refusal behavior) is a feature to design with, not fight: route refusals to humans; never auto-retry refusals in a loop.
Stakeholder Communication & Lifecycle (14%) — the differentiator
The consulting layer unique to Professional:
- Executive framing: outcomes and risk posture, not architecture diagrams; cost as unit economics (per ticket/document), not token trivia.
- Expectation setting: agentic systems are probabilistic — communicate accuracy as measured ranges with eval evidence, never "it works."
- Lifecycle: pilot → eval-gated expansion → production with monitoring → renewal cadence (models, prompts, evals age).
- Change management: champions, training paths (Academy courses per role), and a feedback loop from users into the eval set.
Developer Productivity & Enablement (7%)
Rolling Claude Code out: shared CLAUDE.md + settings in repos, curated skill/command libraries, permission baselines by team, hooks for org guardrails, usage/cost dashboards, and an internal enablement path (101 → in-action → building agents).
🎯 Professional's flavor in one line: Foundations asks "what's the right mechanism?"; Professional asks "what's the right program?" — same principles, org-scale blast radius.