Lesson 2.2
Verification Habits and Source Checking
A proportionate routine for checking Claude's output before it leaves your hands, from a thirty-second scan to a full source audit.
Knowing that Claude can be wrong is not enough; you need a habit that catches the errors without turning every task into an audit. The exam frames this as proportionality: the depth of your check should match the consequences of being wrong. A lunch invitation needs a glance. A regulatory filing needs every claim traced to a source.
The proportionality principle
Ask two questions of any output. Who will rely on it, and what happens if a detail is wrong? Then pick a tier.
| Tier | Typical output | What you do |
|---|---|---|
| Glance | Internal notes, brainstorm lists, first drafts you will heavily edit | Read once for sense; fix anything obviously off |
| Scan | Emails to colleagues, meeting summaries, internal slides | Check names, dates, numbers and commitments against your own knowledge |
| Check | Customer communications, reports to management, anything with figures | Verify each factual claim against a source; re-run key numbers |
| Audit | Legal, financial, regulatory, medical, public statements | Trace every claim to a primary source; second-person review; document what was checked |
The exam expects you to place a scenario in the right tier and to recognise that "I read it and it sounded right" is a glance, not a check.
The thirty-second scan
Even at the lowest tiers, a fast scan catches most problems. Train yourself to look for:
- every proper noun (is this person, company or product real and spelt correctly?);
- every number (does it match the source, and is the unit right?);
- every date and deadline;
- every commitment or promise ("we will", "guaranteed", "by Friday");
- every claim introduced with "studies show", "according to", "it is well known".
Each of these is either verifiable or should be removed.
Checking against sources
For the Check and Audit tiers, the rule is simple: a claim is verified when you have seen it in a source you trust, not when Claude says it is true and not when a second AI answer agrees. Asking Claude "are you sure?" is a weak check; it may apologise and change a correct answer, or confirm a wrong one. Instead:
- Open the source. If Claude cites a document you uploaded, find the passage. If it cites a web page, visit it.
- Prefer primary sources (the regulation itself, the company's own filing, the original dataset) over secondary summaries.
- For numbers, recompute at least one figure by hand or in a spreadsheet.
- For quotations, find the original and compare word for word.
- For anything about your organisation, check with the system of record (HR system, CRM, finance ledger) rather than memory.
Web search inside Claude helps here because it returns links you can open. Treat the link as the thing to verify, not Claude's paraphrase of it.
Using Claude to help verify
Claude can speed verification without being the final arbiter.
List every factual claim in the draft above as a numbered table with
three columns: the claim, where it came from (my uploaded file, web
search, or your general knowledge), and how confident you are. Do not
rewrite the draft.
This produces a checklist. Anything tagged "general knowledge" with a precise figure goes to the top of your verification pile. You can also ask Claude to find the exact passage in an uploaded document that supports a statement, which is faster than searching manually, and then read the passage yourself.
Scenario: the tender response
A bid manager uses Claude to draft a response to a public tender. The draft includes the company's ISO certification numbers, three case-study metrics and a claim about staff headcount. The bid manager runs a Check: certification numbers against the certificates on file (one digit was wrong); case-study metrics against the original project reports (one was from a different project); headcount against HR (correct). Twenty minutes of checking prevented two errors in a legally binding submission. Note that the errors were not exotic; they were precisely the kinds of detail the scan list targets.
Building the habit
- Decide the tier before you read the output, not after.
- Keep a short personal checklist (names, numbers, dates, promises, sources) and apply it every time.
- Never forward Claude's output in the same breath you received it; a deliberate pause is the cheapest control there is.
- Record what you checked for Audit-tier work, so a colleague or auditor can see your trail.
- Feed corrections back into your prompts and Project instructions so the same error does not recur.
Exam focus
- Verification depth should be proportionate to who relies on the output and the cost of error; identify the tier first.
- "Reading it and it sounded right" is a glance, not a check; asking Claude "are you sure?" is not verification.
- A claim is verified only when traced to a trusted source, preferably primary; Claude's paraphrase is not the source.
- Claude can accelerate verification by listing and labelling claims or locating supporting passages, but the human confirms.
- Names, numbers, dates, commitments and attributed claims are the five items to check on every scan.
Check yourself
5 questions
1.Which verification approach is the most robust for a factual claim in Claude's output?
2.A market analyst needs the current line-up of a competitor's products, which changed recently. How should Claude be used so the answer can be verified?
3.A team wants a verification policy that is proportionate rather than exhaustive. Which output deserves the most rigorous checking?
4.An analyst routinely asks Claude 'What are your sources?' after an answer. Which statement about this habit is accurate?
5.With web search on, Claude gives a consultant a market-size figure and links to a blog post. The blog post attributes the figure to unnamed 'industry research'. What is the correct verification step?