Domain 2 · 21% of the exam
Output Evaluation and Validation
The heaviest domain. Judge whether outputs answer the question, spot hallucinations and bias, verify against a source of record, know when a human must review, and shape outputs for the audience.
Task statements
- 2.1
Accuracy and Completeness
Before polishing, check two things: did the output answer the question that was asked, and did it leave out anything material? A fluent response that quietly skips a caveat, an exception, or half of the request is still a failed response.
- 2.2
Hallucinations, Inconsistencies, and Bias
Claude can produce text that sounds right but is not grounded in your source or in fact. Learn the red flags: invented numbers and citations, detail beyond what the source contains, internal contradictions, uniform confidence on every claim, and one-sided framing.
- 2.3
Fact-Checking and Validation
Validation means checking a claim against something independent of the model: the source of record. Self-reported confidence is not validation, and asking Claude to double-check its own answer is weaker than looking at the original document, system, or authority.
- 2.4
When Human Review Is Required
Some outputs should never leave your hands without a person reviewing them. The more external, consequential, or regulated the use, the stronger the need for human review. Claude drafts and assists; accountable people decide.
- 2.5
Edit and Adapt for Audience
A correct answer can still be wrong for the reader. Adapting tone, length, and level of detail for a specific audience is a normal part of using Claude, and comparing drafts side by side helps you choose deliberately.
- 2.6
Output Formats
The shape of an answer matters. Short conversational replies suit inline responses, standalone documents and reusable pieces suit Artifacts, and comparisons or lists suit tables and structured layouts. Pick the format that fits how the result will be used.