An evidence-led reading reverses the order. It begins with what is on the record: who said what, in which document or broadcast, at what time, and who else has said the same. The frame comes last, and it is held loosely.

Start with the claim, not the story

A story is a bundle of claims. “Talks resumed on Tuesday” is one claim. “Talks resumed after pressure from a neighbouring state” is two: the first can be dated and sourced, the second is an attribution of motive that needs its own support. Breaking a dispatch into separate claims is the first discipline, because each claim can then be traced, scored and challenged on its own terms.

Trace every claim to its origin

For each claim, ask where it first appeared. Many reports that look independent are downstream of one statement, one wire item or one anonymous briefing. Tracing to the origin shows how many separate pieces of reporting stand behind a claim, which is usually fewer than the number of headlines.

• Record the first appearance of the claim, with date and time in one fixed time zone.
• Note whether the origin is a primary document, a named person, an unnamed person or an earlier report.
• Mark every later report as a repeat of the origin or as new reporting.

Show the score and what went into it

Confidence is useful only when the reader can see how it was reached. A bare number invites trust or distrust and teaches nothing. A score that lists its inputs, such as the reliability of the source, the amount of independent corroboration, how recent the reporting is and whether primary documents back it, can be argued with. That is the point. A reader who disagrees can see which input they would weigh differently.

Keep description and conclusion apart

The record can say that three outlets reported a figure, that two of them cite the same ministry statement, and that the statement itself has not been published. It does not need to say whether the figure is right. Readers who must decide, whether to brief a minister, file a report or change a plan, are better served by an exact account of the state of the evidence than by a verdict they cannot inspect.

Describe what the evidence shows. Leave the conclusion to the person who has to act on it.

A short checklist

1. Split the story into claims.
2. Find the first appearance of each claim.
3. Count independent origins, not headlines.
4. Score with named inputs and show them.
5. Date everything, and keep corrections visible.

None of this makes an analyst faster at producing an opinion. It makes the opinion easier to defend, easier to correct and easier for someone else to pick up. That is the standard we build 7Analytica to support.

What a label hides

A single outlet produces many kinds of material: wire copy, original reporting, opinion, translated statements and syndicated pieces. Its record on each differs. Treating the outlet as one thing averages those differences away. The reader learns the outlet’s brand but not how likely this particular claim is to survive checking.

What can be measured

Reliability can be approached as a track record. Over a defined period, how often did claims from this source hold up when independent reporting or primary documents arrived? How often were claims later corrected, and how quickly? How often did the source break a story that others then confirmed? These are countable, and the counts can be shown.

• Corroboration rate: the share of claims later supported by an independent origin.
• Correction behaviour: how often and how visibly the source amends its own reporting.
• Primary-document use: how often claims point to a document a reader can open.
• Ownership and funding: who stands behind the outlet, stated plainly and dated.

Why the number must stay provisional

A measured reliability score is an estimate from a finite sample. A source with ten checked claims and a source with a thousand should not share a level of certainty even if both score 80. Showing the sample size beside the score is part of the measurement, not a footnote to it. So is the date of the last update, because a record from two years ago says little about a newsroom that has since changed ownership.

A score without its sample size and its date is a label wearing a number.

Using it without leaning on it

Source reliability is one input to confidence in a claim, not a substitute for the claim’s own evidence. A strong source can be wrong and a weak one can be right. The sound use of the measure is as a starting weight that the evidence for the specific claim can raise or lower, and that the reader can see and adjust.

That is how the source ledger in 7Analytica is built: each source carries its record, its sample and its last review date, and every score that depends on it links back to that entry.

The reader owns the conclusion

An analyst who briefs a decision-maker is accountable for the judgement, including its reasoning. A tool that hands over a verdict cannot be questioned in a meeting, because there is nothing behind it to point to. The analyst is left either repeating a conclusion they cannot defend or redoing the work to find out whether it holds.

Verdicts hide uncertainty

Most geopolitical claims are neither true nor false in a clean sense. They are partly supported, supported only by one origin, accurate in fact but misleading in framing, or too early to call. Collapsing that range into a label throws away exactly the information a careful reader needs. A claim that is “likely” on thin evidence and a claim that is “likely” on strong evidence look identical once they have been rounded.

What we do instead

• Describe the state of the evidence in plain terms: what is reported, by whom, with what support.
• Show the inputs behind every score so that a reader can weigh them differently.
• Separate what is established from what is attributed and from what is inferred.
• Leave a clear place for the analyst’s own assessment, kept distinct from ours.

A calm voice is part of this

Declining to issue verdicts has a consequence for tone. The writing stays level. There are no alarms where the evidence is thin and no reassurance where it is not. Where something is uncertain, we say what would settle it.

We describe. You decide.

This is slower than a one-line answer. It is also what lets a team trust the record enough to build on it.

Why a silent change matters

Analysts cite what they read. If the page has since changed, the citation points at a text that no longer exists. A colleague who checks the link a week later sees a different claim and may conclude the analyst misquoted. Worse, a decision may have been taken on wording that the publisher itself later withdrew.

Three kinds of change

1. Marked corrections and updates, where the publisher says what changed and when. These are healthy and should be easy to find.
2. Unmarked refinements: typo fixes, formatting and clarifications that do not change the claim. These can be listed without comment.
3. Unmarked substantive changes: a figure, a name, an attribution or the thrust of a headline altered without notice. These are the ones a record must keep.

What keeping the record involves

The practice is simple to describe. Capture each version of a cited page with the time it was seen. Compare versions word by word. Classify the difference. Keep the earlier text, and link the dated capture from any analysis that relied on it. Nothing here accuses anyone. Plenty of edits are honest and improve the piece. The aim is that a reader can see what the page said when it was relied upon.

The record keeps what was said at the time it was said.

Edit history as a signal

A pattern of unmarked substantive edits is itself information about a source, and it feeds the source’s measured record. A single edit says little. Many, in one direction, on sensitive topics, say something that a reader ought to be able to weigh.

In 7Analytica this lives in the forensic record for each dispatch: dated captures, word-level differences and a plain classification, so that a change never goes unseen simply because nobody was watching at the moment it happened.

Describe the choice, not the motive

It is easy to slide from noticing a word to accusing the writer of an agenda. That step requires evidence about the writer that the text alone cannot supply. A more careful practice stays with what is visible: this report uses a verb that assigns responsibility to one party, while that report uses a passive construction that leaves the actor unnamed. The reader can draw their own inference.

What can be compared

• Who is the grammatical subject of each action, and who is left out of the sentence.
• Whether numbers are given as counts, as percentages or as vague quantities, and against what baseline.
• Which side’s statements are quoted directly and which are paraphrased.
• Intensity of vocabulary: neutral, hedged or charged, measured against how other outlets describe the same event.

Why comparison needs a baseline

A word is charged or neutral only relative to how else the event could have been described. Reading one article tells you little. Reading ten reports of the same event, side by side, shows where the language diverges and by how much. The divergence is the finding; the individual words are only its evidence.

Framing is measured by comparison. A single report has no baseline.

Keeping the reading honest

Linguistic analysis can mislead if it is presented as certainty. Counts of loaded terms depend on a word list someone chose. Translation can introduce or remove charge. The method should show its word lists, its sample and its limits, and it should be used to prompt closer reading rather than to settle it.

The linguistic audit in 7Analytica follows that rule: it marks the passages that differ, shows how, and stops short of saying why.

Independent origins

Corroboration means that separate pieces of evidence point the same way. Two reporters who each speak to different eyewitnesses are independent. Two outlets that both rewrite one agency item are not. The test is simple to state: if the first source were wrong, would this report also be wrong? If yes, it adds repetition and no confirmation.

How to count it

1. Find the earliest appearance of the claim.
2. For each later report, decide whether it shows its own sourcing or only cites others.
3. Group reports that share an origin, and count each group once.
4. Report both numbers: how many outlets carry the claim and how many independent origins stand behind it.

The two numbers will often differ widely. A claim carried by thirty outlets may rest on two origins. Showing that gap plainly corrects the sense of certainty that volume produces.

What counts as an origin

A primary document is an origin. So is a named official speaking on the record, a dataset with a known publisher, or a reporter’s first-hand observation. An anonymous briefing is an origin of weaker standing, and we say so. Where the chain of sourcing cannot be followed back, the claim is recorded with that fact stated rather than quietly assigned a source.

Count origins, not headlines.

Why this matters for decisions

Decisions made under time pressure lean on how much confirmation a claim seems to have. Counting independent origins gives that sense a basis. It also shows the quieter case: a claim with a single strong origin and few repeats may deserve more weight than one that is everywhere but rests on nothing firm.