Domain 2 — Output Evaluation and Validation (21%)
21% of the exam and the heaviest domain by some margin, roughly 13 questions. If you study one domain properly, make it this one.
2.1 The three questions to ask of any output
- Is it accurate? Are the facts, figures, names, dates and citations correct?
- Is it complete? Did it answer the whole question, including the awkward part?
- Is it appropriate? Right tone, right audience, right length, and safe to send given who will read it?
An output can pass one and fail another. A flawless, beautifully written summary that quietly omits the one finding your director cares about has failed on completeness. A factually perfect response written in the wrong register for the recipient has failed on appropriateness.
2.2 Recognising a hallucination
A hallucination is a confident, plausible, wrong statement. It does not look different from a correct statement — that is precisely the problem, and why "it sounded confident" is never a verification method.
Learn where the risk concentrates:
- Specific numbers that were not in material you supplied — market sizes, percentages, growth rates.
- Citations, references, case names and URLs, which can be fabricated in exactly the right format.
- Recent events after the model's training cutoff, unless it searched for them.
- Niche or internal specifics — your company's process, a small vendor's pricing, a local regulation.
- Quotations attributed to real people.
Conversely, low-risk output includes rewriting text you supplied, reformatting, summarising an attached document, and general explanation of well-established concepts.
2.3 How to reduce the risk before it happens
Verification is cheaper if the prompt was built to be checkable. The techniques, in the order they earn their keep:
| Technique | What it does |
|---|---|
| Give explicit permission to say "I don't know" | Removes the pressure to fill a gap with something plausible. One of the highest-value single sentences you can add. |
| Restrict to supplied material | "Use only the attached documents; do not use general knowledge." Stops gaps being filled from memory. |
| Ask for direct quotes first | Have Claude pull the exact supporting passages before analysing, so the analysis is anchored to real text. |
| Require citations per claim | Makes the output auditable — you can check the claims that matter instead of re-researching everything. |
| Ask it to re-check its own draft | "For each claim, find a supporting quote; if there isn't one, remove the claim." Catches a useful share of errors. |
| Ask the same question twice | Where answers diverge between runs, you have found the uncertain parts. |
Important — these reduce the error rate; they do not eliminate it. For anything high-stakes, a human still validates. An exam option claiming a technique makes output "reliable enough to skip review" is wrong.
2.4 Deciding how much checking is enough
This is the judgement the domain is really testing. Scale the check to the consequence of being wrong.
| Stakes | Examples | Appropriate check |
|---|---|---|
| Low | Brainstorm, first-draft internal notes, rephrasing your own text | Read it. Move on. |
| Medium | Internal report, customer email, meeting summary shared with a team | Verify every specific claim and figure; check nothing important was omitted. |
| High | Published content, financial figures, anything customer-facing at scale, anything a decision rests on | Verify against primary sources and have a second person review. |
| Regulated or consequential to an individual | Legal, medical, financial advice, hiring, lending, insurance, academic assessment, journalism | A qualified professional reviews and owns the output, and AI involvement is disclosed. |
2.5 Detecting bias and one-sidedness
Beyond factual error, check whether the output quietly took a position: a "balanced" comparison that argues for one option, a candidate summary that leans on characteristics irrelevant to the job, a market analysis reflecting the framing of the sources you supplied. Ask explicitly for the counter-argument, or for the strongest case against the recommendation, and see whether it holds up.
2.6 Refining rather than rewriting
Once an output is nearly right, the efficient path is targeted correction: keep the structure and fix the section, or keep the content and change the register. Starting over discards the parts that were working and usually reintroduces problems you had already fixed.
← Previous: Domain 1 — Prompting and Task Execution | Next: Domain 3 — Product and Model Selection →
Explore Other Certification Study Guides
Switch to another exam guide to prepare for your next Claude certification:
Claude Certified Architect — Foundations
Core architectural principles: agentic loops & orchestration, tool design & MCP, Claude Code workflows, prompt engineering, and context reliability.
Claude Certified Architect — Professional
Advanced enterprise architecture: solution design, system integration, evaluation frameworks, governance, safety & developer enablement.
Claude Certified Developer — Foundations
Build software on Claude: API integration, model selection, tools & MCP servers, agents vs workflows, security, safety & debugging.