01A claim is extracted wrongly
A sentence can be split badly, a quotation read as the writer's own claim, or satire read as fact. Check the claim against its passage before relying on it.
Where to look
The passage printed beside the claim
AI And its limits
A model reads. A person judges.
7Analytica uses AI models to do the slow reading of news reports. They do not decide what is true, and they go wrong in specific ways that this page lists. In plain words: software does the first read of the news, and a person decides what it means.
Read The short version
A model reads. A person judges.
Models help read the reporting. They do not decide what is true, and they are wrong in specific, listable ways. This page says which parts are model-assisted, how you can check each one and how to tell us about an error.
Models do the slow reading: splitting reports into claims, weighing the wording, drafting summaries. Everything they produce is meant to be traceable to a passage, a formula or a source you can open.
7Analytica follows one rule from the editorial principles: describe, never conclude. The page you are reading describes the design. Read it alongside the coverage page.
No verdicts Reviewed 1 October 2026
01 Who makes what
Eight parts. Three makers.
A model proposes some parts, a fixed rule makes others and people set the rest. Whoever made a part, you can check it.
Read Each part, who makes it, and how you check it
What is model-made, and how you can check it
Eight parts of a record, who or what makes each, and the check open to you. Model-assisted means a model proposes and a person can verify. Rule-based means a fixed procedure.
| Part | Made by | How to check it |
|---|---|---|
| Claims pulled from a report | Model-assisted | Each claim links to the passage it came from. Read the passage and see whether the claim says what it says. |
| Origin and repeat marking | Rule-based | Compares first-seen times and the text of linked reports. Open the chain and read the timestamps. |
| Source reliability reading | Set by people | Comes from a published track-record table with the outlet's ownership shown. In the demo these values are samples. |
| Trust score | Rule-based | A formula over five inputs, each printed with its weight. The methodology calculator lets you redo the arithmetic. |
| Bias and framing reading | Model-assisted | Lists the words and phrases that moved it. It reads wording, not intent, and you can disagree with any term. |
| Summaries and brief text | Model-assisted | Written from the record and labelled as such. Compare the sentences with the dossier they summarise. |
| Ask 7Analytica Agent answers | Model-assisted | Should point to records. Open the record it cites. If it cites none, treat the answer as unsupported. |
| Seal and edit history | Rule-based | Not written by a model. A seal fixes content and time, and any later change appears as a new entry. |
02 Trace
Nothing a model writes floats free.
Every claim links to the passage it came from. The origin is the first report in the record, and a repeat is a copy of it. A sealed dossier fixes what you saw.
Read How provenance is kept
Nothing floats free
A sentence a model writes is never shown without the thing it came from. Provenance is how you test the model, so it is part of the record, not an extra.
Chain of custody, one claim
Sample values
Step 1: Each claim keeps its passage
A claim links to the report and the passage it was drawn from, with the time the report was first seen.
Step 2: Origins and repeats are marked
The first report, or one with its own sourcing, is the origin. Copies are repeats. counts independent origins only.
Step 3: Gaps stay visible
Where a source was edited, removed or never collected, the chain shows a gap. It is not filled in with a guess.
Step 4: The seal fixes what you saw
Sealing records the content and the time. Anything that changes afterwards is a new entry.
03 Failure modes
Where it goes wrong, listed.
Each entry says where in a record to look. The list is not complete, and it is here so you can catch a mistake before you rely on it.
9Known ways it goes wrong
Read Where to look when a record seems off
Ways it goes wrong
Listed so that you look in the right place when a record seems off. Each entry names where to look. The list is not complete.
Failure map, a record read top to bottom
Numbers match the ledger below
- 01Claim
- 02Origin and repeats
- 03Wording reading
- 04Summary
- 05Agent answer
- 06Coverage
- 07Trust score
02The origin is missed
If the first report is behind a paywall, offline or not collected, a relay can be marked as the origin. Origin means first seen in the record, not first in the world.
Where to look
The first-seen times in the chain
03Paraphrase hides a repeat, or a coincidence looks like one
Repeat detection compares wording and timing. A heavily reworded copy can pass as independent, and two honest reports on one event can look related.
Where to look
The linked reports, read side by side
04Translation changes the framing
Wording differences can come from translation rather than from the outlet. The demo reads English only, which narrows this risk and also narrows what it sees.
Where to look
The source's own wording, before any translation
05Framing is read as intent
A bias reading describes how far wording leans. It says nothing about why the writer chose it.
Where to look
The word list under the reading
06A summary sounds surer than the record
Fluent text can carry more confidence than the evidence does. The score and its inputs, not the prose, are the thing to check.
Where to look
The score and its inputs, not the prose
07The agent over-reads or invents
A language model can state something the records do not say. Treat an answer with no cited record as a lead, not a finding.
Where to look
The record the answer cites
08A gap is read as calm
A with few outlets or a fetching gap can look quiet. The record shows silences as signals and thin zones as thin, but only where it knows about them.
Where to look
The coverage page, and how many outlets a zone has
09A score ends the checking
People trust a number. A high means the evidence in the record stands up well, not that a claim is true.
Where to look
The five inputs and their weights
04 Corrections
We correct in the open.
A correction says what was wrong, what is right and when it changed. A sealed dossier is never rewritten: the fix appears as a new dated entry beside it.
Read Our corrections policy
We correct in the open
7Analytica keeps the edit history of what it reads. It holds itself to the same standard. Changes are logged in the changelog.
Correction notice
SampleAs published, 3 Aug 2026
“The figure doubled between March and June.”
Corrected
“The figure rose by a third between March and June.”
Why: the first report read a quarterly total as a monthly one.
Corrected 14 Aug 2026
Sealed dossier kept as issued
Correction log
- 14 Aug 2026RecordClaim text corrected. The notice above.
- 14 Aug 2026MethodAn input weight restated. The reason is in the changelog.
- 2 Sep 2026SiteWording fixed on a page. Its review date moved.
We say what was wrong
A correction states what was wrong, what is right and when it changed. It is not a quiet edit.
Sealed dossiers are not rewritten
A stays as issued. A fix appears as a new, dated entry beside it.
Method changes are recorded
If a score's inputs or weights change, the change is recorded in the changelog with the reason.
Errors on this website
A mistake in the text of this site is corrected on the page, and the page's review date is updated.
Next
Found a mistake? Tell us.
A model error, a wrong score, a misread claim or a mistake on this site. We read every report.
Read What to send, and what happens next
Found a mistake? Tell us
A model error, a wrong score, a misread claim or a mistake on this site. We read every report.
Error report slip
Date ______ No. ______
What to send
The page or record, what it says, what you think is right and how you know. A link or a screenshot helps.
What happens next
We read it and reply to the address it came from. We say what we found and, if we change something, what changed.
Credit
If you want to be named once it is fixed, say how.
bhuvanchandra@7analytica.com
Opens your email, subject filled in
Security issues
A vulnerability is a different route. Use the disclosure steps in the trust centre.
Read Questions
Questions about the models
Answer
1 of 4
Does 7Analytica use AI?
Yes. Models help pull claims out of reports, read the wording for , draft summaries and answer questions in the agent. Origin marking, score arithmetic, sealing and edit history are rules, not models.
Answer
2 of 4
Can I check how a score was made?
Yes. Every score prints its inputs and weights, and the methodology page has a calculator that rebuilds a sample score line by line. A score is an estimate of how well the evidence stands, not a finding.
Answer
3 of 4
Does 7Analytica decide what is true?
No. It publishes no verdicts. It describes what was said, where and when, and the evidence, and leaves the conclusion to you.
Answer
4 of 4
Are these models running today?
This page describes how each part is designed to work and where it can fail. See the coverage page for what is included.
Check a record yourself. Open the demo, pick a score and read its inputs. The point of this page is that you can. Open the methodology calculator or try the live demo.