CCA-Foundations exam guide update — diff analysis (2026-07-02)

Founder sent the exam guide fresh off a maintenance window, flagged as changed. Read all 37 pages. Bottom line: domain weights/structure are unchanged vs. the vault's most current baseline ([[2026-05-27-claude-certified-architect-foundations-study-plan]], itself built from the actual Anthropic EAP guide). What's new is mostly resolution of previously-unknown exam mechanics and a large amount of granular, official prep material that wasn't in the vault before — plus one genuinely actionable scope cut (streaming) and one stale mislabeling in cert-progress.md that this guide exposes.

The document is versioned — "Version 0.2, Last Updated: June 30 2026" — confirming this is a real update, not the same artifact re-sent.

What did NOT change (confirmed stable)

Domain weights and numbering are identical to the 05-27 study-plan baseline:

No new domain, no dropped domain, no reweighting. Prior harness-mechanics knowledge (PostToolUse can modify tool output before the model sees it, hooks > prompt instructions for deterministic guarantees, subagents don't inherit parent context, 4-5 tools per subagent not 18+, tool_choice auto/any/forced) is all explicitly confirmed in the new task statements — nothing in the cheat sheet was contradicted.

Important caveat on "no change": this is true relative to the vault's current baseline (05-27 doc). It is emphatically NOT true relative to the older anthropic-certification-study-guide.md (dated 2026-04-08), which used a completely different 5-domain structure (D1 Agentic 27% / D2 Claude Code Config 20-23% / D3 API & SDK Integration 20% / D4 MCP Implementation 18% / D5 Production Deployment 12%). That document is now clearly stale and superseded — it was already superseded by 05-27, but this guide is the final confirmation. Recommend archiving or clearly marking anthropic-certification-study-guide.md as superseded so it doesn't get pulled into study sessions by mistake.

What's genuinely NEW

1. Exam mechanics — previously unknown, now fully specified

The 05-27 doc explicitly flagged "duration / # questions / cost / registration / EAP-window: not in this guide." All now resolved:

The 720 passing bar and 60-question count were already inferred from practice exams (cert-progress.md), but this is the first official confirmation, plus new info: $125 fee, 12-month validity, delivery options. No more "confirm the real cut score" hedge needed — 720/1000 is now sourced from the primary document, not founder recollection.

2. Full granular task statements — replaces vague domain scope

The 05-27 doc had one-line paraphrased "scope" per domain. The new guide has 27 numbered task statements (D1: 1.1–1.7, D2: 2.1–2.5, D3: 3.1–3.6, D4: 4.1–4.6, D5: 5.1–5.6), each with explicit "Knowledge of" and "Skills in" bullets. This is the actual blueprint — good for a final self-quiz pass ("can I state the knowledge/skill bullet for each of the 27 statements cold").

3. Six named scenarios + primary-domain mapping

Customer Support Resolution Agent, Code Generation with Claude Code, Multi-Agent Research System, Developer Productivity with Claude, Claude Code for CI, Structured Data Extraction — each tagged with which domains it primarily draws from. Useful for exam-day pacing intuition (4 of these 6 will appear).

4. Twelve worked sample questions with full explanations

Drawn from 4 of the 6 scenarios (Customer Support ×3, Code Gen ×3, Multi-Agent Research ×3, CI ×3), explicitly stated as "drawn from the practice test." This is real calibration material for the exact reasoning style graded on the real exam — see weak-axis section below, this is the highest-value new artifact for his two known gaps.

5. Explicit In-Scope / Out-of-Scope lists — the most actionable new content

Out-of-scope, confirmed will NOT appear: fine-tuning/training custom models, API auth/billing/account management, deep language/framework implementation, MCP server hosting/infra (networking, containers), Claude's internal architecture/training/RLHF, embedding models/vector DB implementation, computer use, vision/image analysis, streaming API implementation or SSE, rate limiting/quotas/pricing calculations, OAuth/key rotation/auth protocol details, cloud-provider-specific configs, performance benchmarking/model comparison metrics, prompt caching implementation details (beyond knowing it exists), token counting/tokenization specifics.

6. Four structured Preparation Exercises (27 numbered steps total)

Hands-on build sequence mapped to specific domains — a ready-made drill plan if he wants project-based reinforcement rather than more practice exams.

Changes touching his two known weak axes

Axis 1 — test-craft "which is BEST" discrimination

The 12 sample questions are a direct hit on this. Several explanations show the exact reasoning pattern his exam-2 miss analysis identified (comprehensive/layered answer beats a clean single mechanism):

• Q12 (14-file PR review) — correct answer is "split into focused per-file passes + separate integration pass," which is literally his exam-2 Q56/Q4 pattern, now codified as Task Statement 4.6 ("Design multi-instance and multi-pass review architectures").
• Q9 (verification round-trips) — worth flagging as a nuance correction to his "most-complete = give it everything" instinct: the correct answer is a scoped verify_fact tool for the 85% simple case while still routing the 15% complex case through the coordinator — NOT giving the synthesis agent full web-search access. "Most complete/robust" does not mean "most access" — it means covering the actual distribution of cases without over-provisioning. Worth adding to the test-craft rule.
• Q2 (tool misrouting) — correct answer is the cheapest highest-leverage fix (better descriptions), not the "more thorough-sounding" routing layer or tool consolidation — a counter-example showing "BEST" isn't always "most machinery," sometimes it's "root cause, minimum effort." Also worth folding into the heuristic: BEST = addresses the actual root cause proportionately, not "the most engineering."

Net: 5 sample questions worth doing as a fresh drill set for this exact weak axis, with two refinements to the existing "most-complete" heuristic (don't conflate "complete" with "maximal access/effort").

Axis 2 — D1 act-vs-surface-to-user judgment

Task Statement 5.2 ("Design effective escalation and ambiguity resolution patterns") is essentially a formalized version of his exam-2 act-vs-ask rubric, with three concrete refinements not previously captured:

• Escalation triggers are explicit customer request for a human, policy exceptions/gaps (not "complex cases" generally), and inability to make progress — his rubric already had "material scope/resource change → surface," this adds the sharper "policy is silent/ambiguous on the specific request → escalate" trigger.
• Multiple ambiguous matches → ask for more identifiers, never select by heuristic. New, specific, not in his existing rubric table.
• Sample Question 3 (55% FCR, escalating easy cases while over-autonomizing hard ones) is a near-exact restatement of his exam-2 Q15/Q17 pattern, with the fix stated as "explicit escalation criteria + few-shot examples" — reinforces that this is a prompt-design fix, not a confidence-score or ML-classifier fix (both distractors in the sample question, mirroring his Q17 "self-modifying params without goal-check" miss).

Net: no new axis, but a materially sharper rubric. Recommend updating the act-vs-ask table in the exam-2 miss analysis with the "policy silent/ambiguous" trigger and the "ask for more identifiers, don't heuristic-guess" rule.

Stale content this update exposes

1. cert-progress.md mislabels a weak area by domain. The entry "D3 API specifics (Message Batches, tool_choice, streaming)" uses the OLD April-8 domain numbering (where D3 = API & SDK Integration). Under the current official numbering (confirmed by this guide), tool_choice and Message Batches live in D4 (Prompt Engineering & Structured Output, Task Statements 4.3 and 4.5) — not D3. Relabel this entry so future /cert-study weak-area tracking tags the correct domain.
2. Streaming event-sequence drilling is now a wasted-effort item. cert-progress.md's queued drill ("streaming event sequence: message_start → content_block_delta → message_stop") targets a topic the guide explicitly places in Out-of-Scope Topics ("Streaming API implementation or server-sent events"). Drop this from the study queue.
3. Extended thinking / adaptive thinking is omitted from the new guide entirely — not in any of the 27 task statements, not in the Appendix "Technologies and Concepts" list, not explicitly in-scope or out-of-scope. The old April-8 doc treated this as core D3 API material with a practice exercise. Given the guide's silence, deprioritize hard — don't spend more study time here beyond what he already knows from daily use. If it appears, it'll likely be a minor distractor, not a tested mechanism.
4. Prefilling ({ token-lock) technique doesn't appear anywhere in the new task statements or Appendix, despite being a real exam-2 miss (Q33) and a documented production technique. Task 4.3's structured-output guidance covers only tool_use/JSON-schema/tool_choice — no mention of prefill as a technique. Possible signal the syllabus has narrowed away from it; can't confirm with high confidence. Keep the concept banked (it's real and correct) but don't over-invest additional study time chasing it as a named exam topic.
5. The entire April-8 anthropic-certification-study-guide.md (model pricing table, cost-modeling exercise, cache-boundary audit, "Production Deployment" domain) is now clearly obsolete under the confirmed current domain structure. Mark it superseded — see note above.

Prep-plan impact

Early-sit plan: still on, lean toward proceeding at the existing pace rather than accelerating further or pulling back. Reasoning:

• No new content area surfaced. The update narrows scope (explicit out-of-scope list removes streaming, vision, computer use, fine-tuning, MCP hosting/infra, pricing/rate-limit calculations, auth/billing, tokenization — several of which the stale April-8 doc had him worried about) rather than adding to it.
• The two known weak axes get sharper drilling material (12 worked sample Qs, refined escalation rubric) but no new axis to learn from scratch.
• Previously-unknown logistics (cost $125, 12-month validity, online-proctored option) remove friction/uncertainty around scheduling — no blocker there.
• Partner-portal gate status in cert-progress.md is still unconfirmed (☐), but is now functionally moot: he has full official guide access regardless of gate state. Still worth a quick explicit confirm before registering, but not a study blocker.
• Exam 2 (2026-06-27) landed borderline at 70% (cut is 720/1000, practice-exam percentage isn't a direct 1:1 mapping to the scaled score, worth remembering). Given no new practice exam has been taken since, recommend one more targeted study pass (the 27-task-statement checklist + the 12 sample questions + the refined act-vs-ask rubric, dropping the now-deprioritized streaming/extended-thinking items) then a third practice exam before locking a sit date. If that lands comfortably above the borderline, sit within 1-2 weeks — that keeps the "pass within first 3 months" escalator (~2026-08-26) very achievable without rushing.

Cheat-sheet sections now stale / needing an update pass

• [[2026-06-14-cca-harness-mechanics-cheatsheet]]: no content is wrong, but it doesn't yet capture the "most-complete ≠ most access" nuance (Q9/Q2 above) or the "policy silent/ambiguous → escalate" + "multiple matches → ask, don't heuristic-guess" rubric refinements from Task 5.2. Recommend a short addendum, not a rewrite.
• [[cert-progress.md]]: the "D3 API specifics" weak-area label needs relabeling to D4 (see stale item 1); the queued streaming drill should be struck (stale item 2).
• [[anthropic-certification-study-guide.md]] (2026-04-08): recommend marking status: superseded in frontmatter and pointing to this note + the 05-27 study plan as current.

Related

• Cached PDF: /tmp/imessage-cache/1783016844-Anthropic_Claude_Certified_Architect_Exam_Guide.pdf
• Durable copy: [[assets/2026-07-02-cca-exam-guide-updated.pdf]]
• [[2026-06-14-cca-harness-mechanics-cheatsheet]] · [[2026-06-27-cca-practice-exam-2-miss-analysis]] · [[cert-progress]] · [[2026-05-27-claude-certified-architect-foundations-study-plan]] · [[anthropic-certification-study-guide]]