Reviewer's Stance
The working posture toward every Claude output: a strong first draft from a brilliant colleague who can be confidently wrong. Fluency and confidence are properties of the writing, not evidence about the facts, and the professional who sends the work owns it.
Exam context: The foundation of the exam's heaviest-weighted domain. Keyed answers check the output against its source before use; distractors send it because it reads well or because Claude sounded sure.
See also: 2.1 Evaluating Outputs for Accuracy and Completeness
Accuracy Pass
The first review question: are the specifics right? Every number, name, date, title, quoted figure and reference gets checked against the source material, ordered by consequence, with the figure a decision rests on checked first.
Exam context: A draft with twenty accurate sentences and one wrong subtotal is a wrong draft if the subtotal is what the meeting decides on. Under time pressure, the keyed answer checks the most costly details first, not the easiest ones.
See also: 2.1 Evaluating Outputs for Accuracy and Completeness
Completeness Pass
The second review question: is anything missing? Checked by walking the source material's own structure, its section list, funds, themes or line items, and confirming each element is represented in the output or knowingly excluded.
Exam context: Omissions leave no trace in the prose, so completeness can never be checked by reading the output alone. Distractors verify one named figure while the stem hints at a dropped section; the keyed answer opens the source.
See also: 2.1 Evaluating Outputs for Accuracy and Completeness
Traceability
The property of an output whose claims can be followed back to their origins. Built by asking Claude to indicate where in the attached material each figure or claim came from, then spot-checking the claims that carry weight.
Exam context: A traceable output is checkable in minutes; an untraceable one gets no benefit of the doubt. Anthropic's prompting guidance backs the habit: for long documents, ask Claude to quote the relevant parts of the source before carrying out the task.
See also: 2.1 Evaluating Outputs for Accuracy and Completeness, 2.3 Fact-Checking and Validation Techniques
Hallucination
Content that is fluent, specific and fabricated: quotes, figures or instructions that look authoritative or sound convincing but are not grounded in fact. An invented detail arrives with precisely the same confidence as a real one.
Exam context: Specificity is not evidence. The spotting question is never "does this sound right?" but "where would this have come from?", and the keyed response verifies against an authoritative source.
See also: 2.2 Identifying Hallucinations, Inconsistencies and Bias
Capability Hallucination
Claude falsely claiming to have done something outside the conversation: sent an email, saved a file elsewhere, produced a working external link. Claude has no access to tools that are not explicitly integrated.
Exam context: A completion claim is a claim to check, not a delivery confirmation. Distractors treat "the email has been sent" as evidence the email exists; the keyed answer verifies that the action actually happened.
See also: 2.2 Identifying Hallucinations, Inconsistencies and Bias
Grounding
The material Claude actually had to work from. Hallucination risk tracks grounding: it is highest for private internal systems Claude was never given documentation for, events after its training data, and precise figures that were not in the source material.
Exam context: Confident numbered instructions for an organisation's own software are the highest-risk shape of all, fluent, specific and built on nothing. The less grounding Claude could possibly have, the less its specificity means.
See also: 2.2 Identifying Hallucinations, Inconsistencies and Bias
Inconsistency
The output disagreeing with itself: a number in the prose that contradicts the table, two sections giving different dates for the same event, a recommendation that contradicts the findings.
Exam context: An internal contradiction is a reliability warning about the whole output, not a typo to patch. The keyed response re-verifies the entire output against its source; fixing the one mismatch you noticed and sending the rest is the tested wrong answer.
See also: 2.2 Identifying Hallucinations, Inconsistencies and Bias
Bias
Skewed framing, assumptions or language, most consequential in outputs about people. It lives in the pattern across descriptions, not in any single sentence: who gets caveats, whose achievements get qualifiers, which groups are framed in different terms.
Exam context: Read across descriptions rather than sentence by sentence, and apply the read-aloud test: would the framing survive being read aloud to the group it describes? The keyed response reworks the pattern; deleting one loaded word leaves the skew intact.
See also: 2.2 Identifying Hallucinations, Inconsistencies and Bias
Independent Validation
Checking a claim against a source that does not share the output's errors: the primary document, the official publication, the regulator's text, the organisation's own records. Independence is the property that makes validation mean anything.
Exam context: Web search is a legitimate means of reaching an authoritative source, but the source does the validating, never the search. Citations exist to be checked, not to decorate.
See also: 2.3 Fact-Checking and Validation Techniques
Circular Verification
Checking a claim against something that inherits the errors of the thing being checked: a second AI summary, a forum post, or Claude's own earlier output in the same conversation.
Exam context: The near-miss source is a favourite distractor: an option that "checks" the draft against the summary Claude produced earlier. It proves nothing, because no independent information enters the check.
See also: 2.3 Fact-Checking and Validation Techniques
Self-Check
Asking Claude to re-check its own output. A legitimate cheap first pass that catches arithmetic slips and internal contradictions, and nothing more: same model, same blind spots.
Exam context: Self-check dressed as verification is this domain's most reliable trap. The guide's own sample rationale is blunt that self-reported confidence is not a reliable accuracy signal, so a confident "yes, it's correct" is never the keyed validation.
See also: 2.3 Fact-Checking and Validation Techniques
Proportionality
Matching validation effort to consequence: the statistic in a public funding bid gets independently verified in two places, the phrasing in an internal ideas list gets a read-through.
Exam context: Tested in both directions. Skimping on the load-bearing claim is the obvious failure, but uniform maximal verification is a real wrong answer too, not a safe one: it spends limited checking time on prose nothing depends on.
See also: 2.3 Fact-Checking and Validation Techniques
Escalation Triggers
The properties that demand qualified human review before an output is used: an external audience; legal, financial or contractual consequence; decisions about people; safety or health implications; irreversibility once sent. Any one is enough.
Exam context: Notice what is absent: how confident the output sounds and how well it reads. Polish is the one property every Claude draft already has, so it never appears among the triggers.
See also: 2.4 When Human Review or Additional Verification Is Required
Qualified Reviewer
The person competent in the output's subject: legal competence for contract language, the accountant for tax figures, the clinician for clinical claims. Anthropic's Usage Policy sets the standard for High-risk AI Recommendations: a qualified person with the training or experience to evaluate the output in that field, and the authority to change it.
Exam context: Seniority and availability are neither necessary nor sufficient. When an option names a reviewer, check what qualifies them, not what ranks them; a busy director skim-reading a lease clause is comfort, not review.
See also: 2.4 When Human Review or Additional Verification Is Required
Over-Escalation
Routing routine, low-stakes drafts through formal review: the committee for a rota note, the solicitor for every internal email. It buries reviewers, stalls routine work and erodes the escalation channel that high-stakes outputs genuinely need.
Exam context: A tested wrong answer, not a cautious right one, because candidates rarely expect caution to be penalised. The skill being certified is triage; the keyed answer lets careful self-review suffice for genuinely low-stakes material.
See also: 2.4 When Human Review or Additional Verification Is Required
Audience Adaptation
Reshaping an accurate output for its named reader: vocabulary down to the reader's level, sentence load lightened, ordering changed so the urgent item leads. The substance, findings, recommendations and costs, must survive the reshaping.
Exam context: Both failure directions are tested: sending reader-hostile output because its content is correct, and simplifying a serious point into invisibility. The keyed answer adapts delivery, then checks the new version against the original for anything lost.
See also: 2.5 Editing, Adapting and Comparing Outputs for the Audience
Criteria-First Comparison
Fixing the judging criteria before reading any version: nothing factual lost, the right register for the named reader, urgent items first. Then reading against the list and defending the choice against the criteria, not the style.
Exam context: Comparison without criteria defaults to the prettiest prose, which is how a fluent-but-thinner version beats a correct one. Fluency is never a criterion, because it is the property both versions already have.
See also: 2.5 Editing, Adapting and Comparing Outputs for the Audience
Generator, Not Judge
The division of labour for variants: Claude produces the versions and articulates how they differ; the human applies the criteria and picks. Asking Claude which of its versions is better outsources the exact judgement the objective certifies.
Exam context: Distractors have Claude recommend its own best version and the user send it. The keyed answer keeps the choosing with the human, for the same reason self-check fails in validation: self-assessment is not a reliable signal.
See also: 2.5 Editing, Adapting and Comparing Outputs for the Audience
Curation
The deliberate step from working material to work product: starting from what the deliverable owes its reader, extracting exactly that from the conversation, and shaping it into its final container. Most of a working conversation is scaffolding; the deliverable owes its reader the decisions, not the journey.
Exam context: Forwarding the whole conversation delegates the curation to a reader who will not do it. The keyed answer extracts what each reader needs into the container that fits their use.
See also: 2.6 Curating Information and Selecting Output Formats
Three Format Homes
The selection rule for output shape, driven by one question: what does the reader do with this? Ask-and-move-on wants an inline answer, compare-and-scan wants a structured table, reuse-and-revise wants an artifact.
Exam context: Format-by-habit options (everything a table, everything an artifact) sound systematic and are wrong in both directions. Feature and model choice inside a format stem is Domain 3 bait; the keyed judgement is always the container the reader receives.
See also: 2.6 Curating Information and Selecting Output Formats
Artifact
Claude's dedicated window for substantial, standalone content: a document, plan or reference page with its own life. It is changed by asking Claude or by editing directly, branched by editing an earlier message, and can be shared as a link when it must travel.
Exam context: The artifact suits content you will edit, iterate on or reuse outside the conversation. The reusable runbook left as loose chat prose gets lost and re-created; the one-line answer wrapped in an artifact gains ceremony for nothing.
See also: 2.6 Curating Information and Selecting Output Formats