01
1. Start from the draft, not from the subject
Print or copy your draft and number every sentence that asserts something. That list, not your knowledge of the subject, is the object of verification. It is usually longer than expected, because assertions hide in subordinate clauses, in adjectives, in captions and in the headline.
Mark each numbered assertion with its type: an event, a quantity, a role, a document's content, a causal claim, or a characterisation. Types matter because each is settled by different evidence and each fails differently. A causal claim and a quantity claim look similar on the page and need entirely different work.
Then mark the three assertions that would cause the most harm if wrong. Those get checked first and most thoroughly. Verification under deadline is triage, and triage requires deciding in advance what matters most rather than working through the draft in order.
02
2. Atomise each claim into checkable units
A single sentence often contains several independent facts. That a service closed in June after demand halved over two years contains a closure, a date, a direction of change, a magnitude, a period and a causal link. Split them. Each unit gets its own line, and the causal link is marked as the weakest.
Atomising converts anxiety into work. It also reveals the units you cannot check, which is far more useful at the drafting stage than an hour before publication. If three of six units are uncheckable, the sentence needs rewriting rather than further searching.
Keep the numbered units. They become the rows of the verification log, the basis of any hedge in the copy and, if a complaint arrives, the evidence that the claim was considered rather than assumed.
03
3. Go to the record that made the fact
Evidence has a hierarchy. The record created by the body that took the action outranks any description of it. For Irish reporting the primary layer usually includes Acts and statutory instruments as published, official releases from the Central Statistics Office, filings in public registers, agendas and minutes of public bodies, court lists and published judgments.
Treat a press release as a signpost. It tells you that a document, dataset or decision exists and it is written to be quoted; it is not the thing itself. Where the underlying record is not published, the Freedom of Information Act 2014 provides a route to records held by public bodies, on the terms and with the exemptions that Act sets out.
Where you cannot reach the primary record, say in the copy what level of evidence you have. A sentence that names its evidence can be weighed by a reader. A sentence that conceals it asks for trust it has not earned.
04
4. Read a statistical release before you use it
Official statistics come with grammar. There is a reference period, a definition of the unit being counted, a note on whether figures are provisional or final, a revision policy, and often a note explaining a break in the series where a definition changed. Every one of those can invert the meaning of the headline figure.
Three specific checks catch most errors. Confirm what is being counted: people, episodes, cases, households or events. Confirm the period and whether you are comparing like with like. Confirm whether the series is seasonally adjusted, and say which version you used if both exist.
Then write the comparison you actually intend to make and check that the release supports it. A release reporting a quarterly change does not support an annual claim, and a national figure does not support a local one, however tempting the sentence is.
05
5. The four number traps
First, percentages against percentage points. A movement from four per cent to six per cent is a rise of two percentage points and a rise of fifty per cent. Both are true, and readers will assume whichever is more dramatic unless you specify.
Second, counts against rates. More cases in a larger area is not a higher incidence. Third, the moving base: a fall of twenty per cent followed by a rise of twenty per cent leaves you below where you started, because the base changed. Fourth, the small denominator: two more out of five is a forty per cent increase and almost never worth a sentence.
The remedy is mechanical. Write the arithmetic out with units attached. Put the base and the period in the same sentence as the percentage. Give the absolute figure beside any proportion. If the number cannot carry its base and period inside one clause, it needs its own sentence.
06
6. Check images and recordings
For any file you did not create, establish four things: where it came from, whether an earlier version exists, whether the depicted place matches independent imagery, and whether the depicted conditions match the record. Signage, shopfronts, vehicle plates, uniforms, language on posters and the direction of shadows all carry checkable detail.
Weather and daylight are the two most useful independent checks available to any reporter, because both are recorded and neither is easy to fake consistently across a whole frame. A file showing heavy rain on a recorded dry afternoon has a wrong date, a wrong location, or is not what it claims to be.
If a file cannot be placed and dated, it is not evidence of anything and should not be published as though it were. Where it is published as an unverified claim, that must be stated beside it rather than in a note further down the page.
07
7. Test independence, not quantity
Two sources are only two sources if they know the thing by separate routes. Two people who attended the same briefing are one source. Two publications citing each other are one source with an extra step. Before counting a second source, ask how this person or record came to know it.
Some claims cannot be settled by any number of sources: internal intention, a private conversation with no record, a statistic nobody collects. Recognise those early. The honest treatment is to report the dispute rather than the fact, naming who says what and on what basis.
Watch for circular confirmation in your own file. A figure you derived from a release, quoted back to you by an interviewee who read the same release, has not been confirmed. Note the origin of every figure in your log so that this cannot happen quietly.
08
8. Keep the verification log
The log is one row per atomised unit: the claim, the evidence sought, what was found, the identifier of the document or timestamp of the recording, the date checked, and the outcome as confirmed, partly confirmed, unverified or refuted. It takes minutes and it is the difference between a documented process and a memory contest.
Use it as a drafting instrument. Every row marked unverified should correspond to a hedge in the text, and every hedge in the text should correspond to a row. Where your sentence is confident and the log says unverified, one of the two must change before publication.
Keep the log with the notes and media, under the same retention rules, and store it so that a confidential source cannot be identified from it. When a query arrives months later, the log answers it in minutes.
09
9. Hedge honestly, or not at all
Hedging is a precision instrument, not a way to publish something you have not established. There is a real difference between a document showing something, a person stating it, a person alleging it, and a reporter being unable to establish it. Choose the phrase that matches the evidence exactly.
Two failures are common. The first is the vanishing hedge, where a careful qualification in paragraph nine does not appear in the headline or the first paragraph that most readers see. The second is the decorative hedge, where words like reportedly are attached to a claim nobody has actually reported, which is a fabricated attribution.
Where a right of reply was sought and not received, state when and by what means. Where a claim failed a check, say so: a checked and refuted claim is often the most useful sentence in the piece, and it is the sentence readers remember as evidence of good faith.
10
10. Decide, and record the decision
At the end of the pass, three outcomes exist for each unit: publish as established, publish with a stated qualification, or do not publish. Make the decision explicitly and write it beside the row. Anything left ambiguous in the log will be resolved by whoever is subbing the copy at speed, which is not where the decision belongs.
Where a decision is finely balanced or the claim is serious, it belongs with an editor rather than the reporter alone, and the criteria for that referral should be written down in advance. Recording who decided and when takes one line and answers most later questions.
Then close the file. Store the log, the notes, the media and the version published, with a retention period, and note anything you would check first if the story returns. Verification improves fastest when the same person sees, in their own records, which kinds of claim they habitually get wrong.
A verification log, filled in
| Unit | Evidence sought | Found | Outcome |
|---|---|---|---|
| Service closed | Operator's own notice and public register entry | Notice dated, register updated | Confirmed |
| Closed in June | Date on the notice, minutes of the deciding meeting | Notice says 12 June; minutes say decision taken in May | Confirmed, with the two dates distinguished |
| Demand fell | Published series with definitions and reference period | Series shows a fall over eight quarters | Confirmed for the stated period only |
| Demand halved | Arithmetic from the same series, base stated | Fall of 46 per cent, not 50 | Refuted as written; figure corrected |
| Closure caused by demand | Statement from the deciding body | No statement; operator declined to comment | Unverified; reported as attribution only |
The arithmetic that catches most number errors
percentage change = (new − old) ÷ old × 100percentage-point change = new percentage − old percentagerate = count ÷ population × chosen multiplierWorked example: a count falling from 1,300 to 702 over eight quarters is a fall of 598, which is 46 per cent of 1,300. It is not a halving, and the difference between 46 and 50 per cent is exactly the kind of error a correction has to carry later.
Worked example: a share rising from 4 per cent to 6 per cent is two percentage points and a rise of fifty per cent. Write whichever you mean and name it, because the reader cannot tell from the number alone.
A two-day verification pass on a contested story
- Day 1, morningNumber every assertion in the draft, mark its type, and identify the three that would cause most harm if wrong.
- Day 1, middayAtomise those three into units and open a log row for each. Request any record that is not published.
- Day 1, afternoonRead the primary records with their definitions and revision notes before touching the numbers.
- Day 1, eveningWrite the arithmetic out with units and bases. Correct the draft where the figures do not support it.
- Day 2, morningTest source independence: how does each person or record know this, and by what separate route.
- Day 2, middayCheck images and recordings for place, date and prior versions. Discard anything that cannot be placed.
- Day 2, afternoonMatch every unresolved log row to a hedge in the copy, and every hedge to a row.
- Day 2, closeRecord the publish, qualify or hold decision for each unit, with who decided and when.
Checklist for a verification pass
- Every assertion in the draft numbered, including those in captions and the headline.
- Claims atomised into units, with causal links marked as the weakest.
- Primary record obtained, or the evidence level stated in the copy.
- Definitions, reference period and revision status read before the figures.
- Arithmetic written out with units, base and period, and absolute figures given.
- Second sources tested for independence rather than counted.
- Files placed and dated, with anything unplaceable discarded.
- Log rows matched to hedges in both directions.
- Publish, qualify or hold decision recorded for each unit with a name or role.
Questions readers ask
How much verification is proportionate?
It scales with the harm an error would cause and the seriousness of the claim. Triage the draft first: the three assertions that would do the most damage if wrong deserve most of the available time.
Can I publish something I could not verify?
Sometimes, if the uncertainty can be described accurately in the same place as the claim and the public interest supports reporting it. What is never acceptable is a confident sentence with the doubt moved out of sight.
Does a verification log have to be a formal document?
No. A five-column table in whatever tool you already use is enough. The value is in the discipline of one row per unit and a recorded outcome, not in the format.