What You Need to Know
Validation means checking an output's claims against something that can confirm them. This objective gives objective 2.1's review its tools, ordered cheapest to heaviest. A judgement layer decides how much checking a given output deserves. A primary document is the original record: the official publication, the regulator's text, the funder's published criteria, the organisation's records.
Take a removals firm. The office has Claude draft a quote letter from the attached survey notes and price list. It states a total price and the council's parking rules for the street. Claude also drafts internal newsletter ideas.
Claude should not be treated as a single source of truth, and high-stakes output deserves careful scrutiny. Claude can produce fluent, specific output that looks correct but is badly mistaken. Specificity alone is no evidence of accuracy.
Cheapest First: The Source You Gave It
For output built from attached material, the first check is the attachment. It is already in the conversation. Each figure, section and claim matches it or does not. This is objective 2.1's habit as a technique, and it costs minutes. The total is checked against the survey notes and price list. Reach for anything heavier only when the given source cannot settle it.
Independent, Authoritative, Primary
External claims need external checking. The attachments cannot confirm the parking rules, which did not come from them. Web search responses include citations "so you can easily verify sources yourself". Users are advised to cross-reference cited sources and use authoritative ones for critical decisions. The Help Center's hallucination article says to review Claude's cited sources, because checking the original catches misinterpretation. Citations exist to be checked.
The standard is the primary document. A second AI summary, a forum post, or Claude's own earlier output is not independent validation. A check that inherits the errors of the thing being checked proves nothing. For the parking rules, that is the council's published page. Web search from within claude.ai is a legitimate way to reach it. The source does the validating, never the search.
Make the Output Traceable
A traceable output shows where each claim came from, so it is checkable. An untraceable one gets no benefit of the doubt. For long documents, Anthropic's prompting guidance says to ask Claude to quote the relevant parts of the source before carrying out its task. For research work, it says to ask Claude to verify information across multiple sources. On the quote letter, ask Claude to show each claim's source and how it reached each conclusion. Then spot-check the weighty ones. When a single figure drives the decision, triangulate it: confirm it in two independent places before it ships.
Proportionality: The Judgement Layer
Validation effort follows consequence. The customer decides on the total, so it gets two independent confirmations. The newsletter ideas get a read-through. Skimping on the consequential claim is the obvious failure. Heavy checking on low-stakes work is also wrong, not safe. Time spent triple-checking newsletter wording nobody acts on is taken from the total.
Asking Claude to re-check its own output has a limited place. Anthropic's prompting guidance says a self-check catches errors, especially in coding and maths. That guidance names one exception: Claude Opus 5 verifies its own work well without being told. A self-check instruction written for earlier models can push it into over-verification. It is a cheap first pass for arithmetic slips and contradictions. It is not independent validation: same model, same blind spots.
In the exam
Questions test the techniques, and when each is proportionate in both directions. The favourite trap is self-check dressed up as verification. An option has the user ask Claude to confirm, re-check or rate its own work, and calls the confident answer validation. The guide's sample rationale says why it fails: self-reported confidence is not a reliable accuracy signal.
Key Concept
When a validation item offers several checking plans, test each against two questions. Does independent information enter the check? Does the depth of checking match what rides on each claim? The keyed answer passes both. The distractors usually fail one: self-check and circular sources fail independence, uniform maximal verification fails proportionality.
Exam Traps
Asking Claude to confirm or re-check its claims and calling that verification
No independent information enters the check, and the same blind spot that produced the error will confirm it. Self-check is a legitimate first pass for slips and contradictions, but the keyed answer always reaches an independent source for claims that matter.
Checking a claim against another AI summary or Claude's own earlier output
Circular verification inherits the errors of the thing being checked. Independence is the property that makes validation mean anything, so the keyed answer goes to the primary document: the official publication, the published criteria, the organisation's records.
Verifying every sentence to the same maximal depth regardless of stakes
Proportionality is the tested skill: uniform heavyweight checking spends limited time on prose nothing depends on and is the over-engineered detour, not the safe choice. The keyed answer concentrates the checking on the claims a decision rests on.
Practice Scenario
A heritage railway volunteer coordinator has Claude draft a grant application. The draft states annual visitor numbers, a percentage increase in school visits, and that the railway meets the funder's published eligibility criteria. She has one evening, too little to verify every line in depth, and nothing may rest on Claude's word alone. What is the right validation plan before submission?
Build Exercise
Build a Validation Plan for One Real Document
What you'll learn
- How to name the authoritative source that settles each claim
- How to sort claims by consequence and assign checking depth
- How to make an output traceable before spot-checking it
- How to see what self-check alone would have missed
- Take a Claude-drafted document from your work containing at least three checkable claims: a figure, an external fact, and a claim about meeting some published requirement
Why: The three claim types need three different validation moves, which is the toolkit in miniature.
You should see: One document with at least three claims underlined and categorised.
- For each claim, name the authoritative source that would settle it: your organisation’s records, the official publication, the published criteria
Why: A claim with no nameable source is untraceable, and untraceable claims get no benefit of the doubt.
You should see: Each claim paired with a named source, or flagged as untraceable.
- Sort the claims by consequence: give the load-bearing figure two independent confirmations, give low-stakes phrasing a read-through, and write one sentence on why each got the depth it did
Why: Proportionality is the judgement layer the exam tests; the sentence forces the reasoning into the open.
You should see: A checking plan where depth visibly varies with stakes.
- Ask Claude to show where each remaining claim came from and how it reached each conclusion, then spot-check the ones that carry weight
Why: Traceability turns a vague read-through into a checkable review and concentrates effort where it pays.
You should see: Claims tied to sources and reasoning, with the weighty ones confirmed.
- Run the plan and record which claims failed, which source caught them, and what a self-check-only plan would have missed
Why: The comparison against self-check makes the independence principle concrete on your own work.
You should see: A short record: failures found, sources that caught them, and the self-check gap.
Sources
- Enable and use web search — Anthropic
- Claude is providing incorrect or misleading responses. What's going on? — Anthropic
- Use research on Claude — Anthropic
- Prompting best practices — Anthropic
- Claude Certified Associate – Foundations Exam Guide — Anthropic