Accuracy
Whether the claims in an output are correct against the source or the facts. Fluent, confident writing can still be inaccurate.
See also: 2.1 Accuracy and Completeness
Completeness
Whether the output answered the whole request and kept the details that matter. Building a checklist from the source before you read the output makes gaps visible.
Exam context: A summary that is true but drops a deadline or exception is a failure of completeness.
See also: 2.1 Accuracy and Completeness
Omission
Leaving out material information, such as a caveat, limit, exception, or risk. Omissions are hard to spot because every sentence that remains can be true.
Exam context: Omission of material caveats counts as a failure. Check what is missing, not only what is stated.
See also: 2.1 Accuracy and Completeness
Hallucination
Output that sounds plausible but is invented or unsupported, such as made-up statistics, citations that do not exist, or details beyond what a source says.
Exam context: Red flags: exact numbers without a source, citations you cannot locate, detail beyond the document, and confident tone throughout.
Higher risk: Specific figures, quotations, names, dates, and references.
First response: Verify against a source you trust rather than asking Claude to confirm itself.
Inconsistency
Contradictions within one output or between runs, such as two different totals for the same quantity. It signals that something needs to be checked.
Bias
Systematic skew in an output, such as one-sided framing, stereotyped assumptions, or unequal treatment of similar cases. It can come from the input material as well as from the model.
Exam context: Look for unbalanced treatment and missing counterarguments, especially in content that affects people.
Uniform Confidence
An even, assured tone applied to every claim regardless of how reliable it is. Claude's tone does not reliably signal which statements are solid.
Exam context: Self-reported confidence is not validation. Never accept "Claude said it was sure" as a check.
Source of Record
The system or document that officially owns a fact, such as the finance system for revenue or the signed contract for terms. Verification means comparing against it.
Exam context: Verify consequential specifics against the source of record, not against another AI output.
See also: 2.3 Fact-Checking and Validation
Fact-Checking
Checking specific claims against reliable sources. A practical sequence is to ask Claude for evidence, verify that evidence in the source, and have an independent person confirm what drives a decision.
See also: 2.3 Fact-Checking and Validation
Human Review
A qualified person checking an output before it is used. It is required for external-facing material, legal, financial, and health content, decisions about people, and regulated workflows.
Exam context: When stakes are high or the audience is external, the answer includes human review. Claude's output is a draft, not a decision.
See also: 2.4 When Human Review Is Required
Audience Adaptation
Editing tone, length, and level of detail to suit the reader while keeping the facts identical. Comparing versions side by side catches facts that changed in the rewrite.
See also: 2.5 Edit and Adapt for Audience
Artifact
A standalone piece of content, such as a document, table, or page, that Claude produces in a dedicated panel beside the chat. It suits deliverables you will reuse, copy, or revise over several turns.
Exam context: Use an Artifact for standalone, reusable, or iterated content. A short conversational answer does not need one.
See also: 2.6 Output Formats
Structured Format
An output shaped as a table, list, or fixed set of headings so it can be scanned and compared. Specifying the format in the prompt gives more reliable results than leaving it open.
Exam context: A comparison across several attributes is best presented in a table.
See also: 2.6 Output Formats