How a score is built, in full.

This paper explains how 7Analytica turns news reports into readings that a person can check. It covers what is measured, how each score is built, what the scores mean, where they fail, and how to challenge one. It is written for analysts, editors, risk and compliance staff, researchers and anyone deciding whether to rely on a reading.

The status matters, so it comes first. 7Analytica today covers one sector, geopolitics and borders, and reads live data from real sources. The worked examples in this paper use invented sample inputs. The outlet profiles in them come from a sample source list. The corroboration and wording inputs in the worked example are set by hand to show the method. No score in this paper describes the real quality of a real outlet.

The live data service is in place. No outside party has validated the weights, and no error margin has been measured. Those are limits of the method as it stands, and section 11 lists them with the others.

Where this paper and a help article on the site differ, the difference is named in the section concerned. The calculator on the methodology page runs the same arithmetic as the worked example here, so you can reproduce every number.

The objects the readings attach to

• A report is one published item from one outlet, captured with the time it was seen.
• A claim is one statement of fact extracted from reports. Many reports can carry the same claim.
• A source is an outlet or a document. It has a longevity, an ownership and a track record.
• A record is one story with its claims, sources, scores and audit. It has a permanent reference number.
• A dossier is a record exported in a fixed form, which can be sealed.

Two of these readings are about the teller and the telling. Trust is about who is saying it. Bias is about how it is said. Keeping them apart is deliberate: a careful outlet can word a story loudly, and a careless one can word it plainly.

A source is scored on four inputs, each from 0 to 100. Three come from the source list: how long the outlet has published, how much its owner has at stake in the telling, and how often its earlier claims held up. The fourth, corroboration, depends on the claim in hand and is covered in section 4.

The ownership input is a judgment. Someone has to decide that a state owner weighs 0.88 and a nonprofit cooperative weighs 0.12, and the number cannot be derived from a formula. The demo shows the weight beside the outlet so you can disagree with it, and the formula lets you see what a different weight would do.

A claim repeated by many outlets can still rest on one statement. Corroboration is only meaningful when the sources are independent, so 7Analytica follows each claim back to its first appearance and gives every report one of three roles.

1. 01Primary: the document or statement itself, when a reader can open it, such as a posted notice.
2. 02Origin: the first report of a claim, or a report with its own sourcing, such as a named witness or its own observation.
3. 03Repeat: a report that relays an earlier one without adding anything. A repeat is kept in the record and does not add to corroboration.

An example from the sample record

The sample record "One closed port, fourteen reports" ingests fourteen reports about a port closure in its first three hours. Counting origins moves the desk from fourteen voices to three: a port agency wire, a shippers' association statement and the harbour authority's own notice, which is the single primary document. The other eleven repeat.

The role counts matter more than the scores. A repeat can score 60 because its outlet has a long record. It still adds no new evidence about the port.

The limit: classification depends on what can be seen. A repeat that hides its source can look like an origin. When origin cannot be established, the record says so and does not guess.

trust = round( (longevity + corroboration + ownership + veracity) / 4 )

Each input moves the score by a quarter of its own change. Raising ownership conflict from 30 to 40 percent lowers the ownership input by 10 and the trust score by 2.5, before rounding. No single input can sink or rescue a score alone, which is also the cost: a source with a poor record can still reach the middle of the range if it is old and well corroborated.

The printed score always carries the four inputs and each input's distance from the average, so a reader can see which one is holding the score up or pulling it down. In the state-owned profile in section 9, longevity sits 33 points above the average and ownership sits 35 below.

The help article "How a score is built" also lists recency and primary documents among the inputs. The methodology calculator uses the four above and does not include those two. Where they differ, the four-input arithmetic is what the calculator runs.

The bias score reads the wording of a report. It does not say a report is false. Four axes are scored from 0 to 100 and averaged with equal weight.

bias = round( ((100 - accountability) + intensity + framing + veracity axis) / 4 )

Tiers and the Perspective Slider

• A bias reading above 48 raises tier 1, Elevated.
• Above 62, tier 2, Critical.
• Above 78, tier 3, Systemic.
• Above 65, the audit offers the Perspective Slider, which sets other outlets' accounts of the same event side by side.

Because the veracity axis is read from the trust score, a source with a poor record cannot read as neutral on bias even when its wording is plain. That is intended: unverified attribution is itself a lean. It is also a coupling between the two scores that you should know about, because it means they are not fully independent.

The audit breaks one record into who, what, when, where, why and how. Each answer carries a confidence reading from 0 to 100 and a mark if sources contest it. There is no overall audit score on purpose. A record can be solid on what and weak on why, and an average would hide that.

In this record the desk can print the closure and the time as established. It must mark the berths as unconfirmed, and it must give both reasons with their origins. The audit says where the record is thin. It does not pick which reason is right.

Why is usually the lowest-confidence axis. Stated motive is shown as framing, not as fact.

Reporting is grouped into eight geographic media zones: North America, Europe, Asia, South America and the Caribbean, Africa, MENA and Gulf, Oceania and Polar, and a digital and gray zones strip. A country sits in one zone. Real framing sometimes straddles two, so the assignment is a simplification.

Each zone has a heat reading from 0 to 100. The help article describes it as a combination of the volume of reporting, how fast it is rising and how far it departs from the zone's own 30-day baseline. The calculator computes a simpler version from sample dispatches: the average alert tier of the zone's items times 26, plus 7 for each item at tier 2 or above, held at 100. Readings of 75 and above are High, 50 to 74 Elevated, and under 50 Low.

• An anomaly is flagged when a reading sits two or more spreads from its baseline. A flag is an invitation to look. It is confirmed only when independent physical telemetry, such as ship tracking or imagery, agrees with the reporting.
• Tier 3 is raised when two or more tier 2 alerts fall within 60 minutes. The rule that raised a tier is printed beside it.
• A silence sentinel alerts on an abnormal drop in coverage inside an active crisis zone, because a sudden quiet can be a signal.

A thin zone is not a quiet zone. Where the sample holds few outlets, a gap in the record is not evidence that nothing happened.

Both profiles use outlet attributes from the sample source list: years published, ownership conflict weight and veracity record. The corroboration, accountability, intensity and framing inputs are set by hand to show the method. They are the presets of the methodology calculator, so you can load either and watch the lines build.

Established wire profile: trust 89, bias 25

State-owned outlet profile: trust 47, bias 66

Reading the two side by side

• The first profile reads trust 89 and bias 25: no tier mark, and the audit would not ask for a second account by default.
• The second reads trust 47 and bias 66: tier 2, Critical, and because bias is above 65 the audit would offer the Perspective Slider.
• Longevity is not what separates them. The second profile's longevity input is 80, close to the first's 93. The gap comes from ownership (12 against 82) and corroboration (42 against 88).
• The distance of each trust input from the average for the second profile is +33 longevity, -5 corroboration, -35 ownership and +7 veracity.

These are sample inputs. Reading them as a verdict on any outlet that shares a profile's attributes would misuse them.

A score does mean

• An estimate of how well the evidence stands, printed with the inputs that produced it.
• A way to compare two accounts under the same method, and to see which input explains the difference.
• A prompt to look. A low trust score or a raised tier tells you where to spend your attention.
• A reading that moves when evidence moves, with the earlier value kept.

A score does not mean

• That a claim is true or false. A low score on a true claim is possible: it says the evidence held is thin.
• A judgment of an outlet's worth, honesty or intent, or of the people who work there.
• A forecast. The impact matrix shows exposure if a claim held. It predicts nothing.
• Advice. Nothing in 7Analytica is investment, legal or political advice.
• A difference between two numbers a few points apart. The weights are equal by choice, they have not been fitted to outside data, and no error margin has been measured.
• A comparable figure across other tools. Another product's score uses other inputs and other weights.

A related limit sits outside the arithmetic. A score is a summary, and any summary invites the reader to stop at the number. 7Analytica prints inputs and keeps the working one click away for that reason. It cannot make a reader open it.

• Every record has a permanent reference number. It is never reused, so a record can always be cited and found again.
• Reports are captured with the time they were seen, de-duplicated and linked to their sources.
• The chain of custody is the recorded path of a claim or document: who held it, when and what changed.
• Edits are entries. The earlier value stays visible, with the time of the change and what moved it.

A seal fixes a record's content and time. Any later change is recorded as a new entry rather than a rewrite, and the sealed version stays as it was. A seal says what the record held at that moment. It does not say the record was right.

In the sample files on the site, the seal block is a stand-in: nothing in them was sealed or signed. How a seal is implemented has not been published, and this paper makes no claim about it beyond the behaviour above.

1. 01Reproduce it. Load the matching inputs in the methodology calculator and confirm the arithmetic. If the arithmetic differs from the printed score, that is an error to report.
2. 02Name the input. Say which of the four trust inputs or four bias axes you think is wrong, what you think it should be and how you know.
3. 03Send it. Write to the address below or use the contact form. Include the record or page, what it says, what you think is right, and a link or a screenshot.
4. 04Expect a person. We read every report and replies to the address it came from. We say what we found and, if we change something, what changed.

We do not state a reply time, because no service level is in place. A change to a score is logged in the changelog with the reason, and the earlier value is kept. A sealed dossier is not rewritten. A correction appears as a new, dated entry beside it.

If you disagree with a weight and not an input, say so. That is a fair challenge to the method itself, and it belongs in the next edition of this paper.

The version and date are on the cover and in the table below. Dates are in UTC. The paper is regenerated from the same source as the website, so the page and both PDFs always carry the same text.

To cite this edition: 7Analytica, "How 7Analytica builds a score", methodology whitepaper, version 1.0, 7analytica.com/methodology/whitepaper.

Changes to the arithmetic are written in the changelog. This paper is reissued with a new version number when the arithmetic or the inputs change, and not for wording alone.