Methodology and responsible use

How the links are made, what they mean, and what they do not mean.

What a link means

A link says that a later coroner's report raised the same or a closely related hazard as an earlier report to the same organisation, often after the organisation stated an action. It does not say why the concern recurred. It is not a finding that any organisation ignored a warning, caused a death or is responsible for anything. Circumstances, evidence and context differ between inquests.

What the labels mean

Chain: A chain is one hazard raised with one organisation in two or more reports, at least six months apart.
Identical hazard: Identical hazard: the later report raises the same specific risk, by the same mechanism, as the earlier one.
Closely related: Closely related: the same specific risk, in a different setting or with a different emphasis.
Confidence: How sure the judge model was that the two concerns are the same hazard, from 0 to 1. Links below 0.70 are never shown. High is 0.85 or more. A chain is only as strong as its weakest link, so the chain shows the lowest confidence of any of its links.
Verified quote: Every quote shown was checked by code to appear word for word in the original report.
No published response found: No response from this organisation to an earlier report was found on judiciary.uk. It may exist elsewhere, or may not have been published.
Raised again after a stated action: The later report came after the organisation's response stated an action. It shows the concern was raised again, not why.
Meaning, wording and judge scores: TF-IDF compares shared words. Embedding compares meaning. Judge is the Nemotron Super verdict that decides whether a link is shown.

Data source and licence

Reports and responses are published by the Courts and Tribunals Judiciary at judiciary.uk.
The material is Crown copyright and is used under the Open Government Licence v3.0.
Every report and response is linked at its original address. Response documents are not copied or rehosted.
The archive was read politely: one request per second, with a contact address in the request header.
This analysis covers 6,454 unique reports. Republished duplicates are counted once.

How it works

Each report's concerns and circumstances sections are separated from the standard Regulation 28 template text, which is removed so it cannot drive similarity.
Receiver names are normalised into organisations, including known renames such as Highways England to National Highways.
Nemotron 3 Nano splits each report's concerns into separate concerns, each with a quote copied from the report.
Within each organisation, concerns from reports at least six months apart are compared with TF-IDF and embedding similarity. Only the closest pairs go forward.
For the earlier report in each pair, Nemotron 3 Nano lists the actions stated in the published response, each with a quote.
Nemotron 3 Super, with low-effort reasoning, judges each pair as identical, closely related, related or different, and gives a confidence.
Pairs judged identical or closely related, with confidence of at least 0.70, are grouped into chains.
For promised actions in chains, Tavily searches the web for evidence of what happened next, and Nemotron 3 Nano classifies what it finds.
Nemotron 3 Ultra drafts a dossier for the strongest chains. A dossier is published only after a person has read it.

Safeguards

Every quote shown is checked by code as an exact substring of the source text, after whitespace and typographic quote normalisation. A link whose quotes cannot be verified is not shown.
Where a model's quote differed only slightly from the source, the source's own words are shown instead.
Model-written sentences are checked against a list of blaming words and withheld if they contain one.
Names of clinicians, staff and family members are redacted, including names next to a role such as "the triage nurse" or "his mother". Coroners are named as public office holders. The deceased are referred to only by the published report title: in quotes, their names are replaced with [the deceased].
There are no league tables or rankings. Lists are sorted by date.
"No published response found" means none was attached to the report on judiciary.uk. A response may exist elsewhere.
"No evidence found" from the web search does not mean an action did not happen.

An invented example, not from any report

As written in a source

As shown on this site

[name redacted], the triage nurse, said [the deceased] waited four hours, and his mother, [name redacted], called the ward twice.

Limitations

Linking is within one organisation. Renamed or merged bodies are linked only where a rename is known.
Some older reports and responses are scanned images. They were read with OCR, which can introduce errors.
Similarity search can miss a recurrence described in very different words.
The judge model can be wrong. Confidence is the model's own estimate, checked against a labelled sample on the evaluation page.

Prior work

PFD Monitor groups coroners' concerns into themes and extracts stated actions. Its responsible-use guidance says it should not be used to decide that an organisation ignored a warning, and we agree.
The Preventable Deaths Tracker by Georgia Richards tracks responses and overdue replies.
PFD Toolkit by Sam Osian provides an open-source loader and thematic analysis. No PFD Toolkit code is used here.

Open data API

The same published, display-safe data is available as JSON, for researchers and other tools. Quotes are verified and redacted exactly as on the site. Results are ordered by date, never ranked. Up to 60 requests a minute.

GET /api/v1/chains?org=National+Highways&q=hard+shoulderRecurrence chains, filtered by organisation name and words. Also limit and offset.
GET /api/v1/chains/c52cded99f756309One chain: its reports, each verified link with both quotes, stated actions and follow-up evidence.

Corrections and contact

If a link, quote or redaction is wrong, please tell us and it will be reviewed and corrected or removed. Open an issue at https://github.com/Safwan10295/Recurance.

What a link means

What the labels mean

Chain

A chain is one hazard raised with one organisation in two or more reports, at least six months apart.

Identical hazard

Identical hazard: the later report raises the same specific risk, by the same mechanism, as the earlier one.

Closely related

Closely related: the same specific risk, in a different setting or with a different emphasis.

Confidence

How sure the judge model was that the two concerns are the same hazard, from 0 to 1. Links below 0.70 are never shown. High is 0.85 or more. A chain is only as strong as its weakest link, so the chain shows the lowest confidence of any of its links.

Verified quote

Every quote shown was checked by code to appear word for word in the original report.

No published response found

No response from this organisation to an earlier report was found on judiciary.uk. It may exist elsewhere, or may not have been published.

Raised again after a stated action

The later report came after the organisation's response stated an action. It shows the concern was raised again, not why.

Meaning, wording and judge scores

TF-IDF compares shared words. Embedding compares meaning. Judge is the Nemotron Super verdict that decides whether a link is shown.

Data source and licence

Reports and responses are published by the Courts and Tribunals Judiciary at judiciary.uk.

The material is Crown copyright and is used under the Open Government Licence v3.0.

Every report and response is linked at its original address. Response documents are not copied or rehosted.

The archive was read politely: one request per second, with a contact address in the request header.

This analysis covers 6,454 unique reports. Republished duplicates are counted once.

How it works

Each report's concerns and circumstances sections are separated from the standard Regulation 28 template text, which is removed so it cannot drive similarity.

Receiver names are normalised into organisations, including known renames such as Highways England to National Highways.

Nemotron 3 Nano splits each report's concerns into separate concerns, each with a quote copied from the report.

Within each organisation, concerns from reports at least six months apart are compared with TF-IDF and embedding similarity. Only the closest pairs go forward.

For the earlier report in each pair, Nemotron 3 Nano lists the actions stated in the published response, each with a quote.

Nemotron 3 Super, with low-effort reasoning, judges each pair as identical, closely related, related or different, and gives a confidence.

Pairs judged identical or closely related, with confidence of at least 0.70, are grouped into chains.

For promised actions in chains, Tavily searches the web for evidence of what happened next, and Nemotron 3 Nano classifies what it finds.

Nemotron 3 Ultra drafts a dossier for the strongest chains. A dossier is published only after a person has read it.

Safeguards

Every quote shown is checked by code as an exact substring of the source text, after whitespace and typographic quote normalisation. A link whose quotes cannot be verified is not shown.

Where a model's quote differed only slightly from the source, the source's own words are shown instead.

Model-written sentences are checked against a list of blaming words and withheld if they contain one.

Names of clinicians, staff and family members are redacted, including names next to a role such as "the triage nurse" or "his mother". Coroners are named as public office holders. The deceased are referred to only by the published report title: in quotes, their names are replaced with [the deceased].

There are no league tables or rankings. Lists are sorted by date.

"No published response found" means none was attached to the report on judiciary.uk. A response may exist elsewhere.

"No evidence found" from the web search does not mean an action did not happen.

An invented example, not from any report

As written in a source

As shown on this site

[name redacted], the triage nurse, said [the deceased] waited four hours, and his mother, [name redacted], called the ward twice.

Limitations

Linking is within one organisation. Renamed or merged bodies are linked only where a rename is known.

Some older reports and responses are scanned images. They were read with OCR, which can introduce errors.

Similarity search can miss a recurrence described in very different words.

The judge model can be wrong. Confidence is the model's own estimate, checked against a labelled sample on the evaluation page.

Prior work

PFD Monitor groups coroners' concerns into themes and extracts stated actions. Its responsible-use guidance says it should not be used to decide that an organisation ignored a warning, and we agree.

The Preventable Deaths Tracker by Georgia Richards tracks responses and overdue replies.

PFD Toolkit by Sam Osian provides an open-source loader and thematic analysis. No PFD Toolkit code is used here.

Open data API

GET /api/v1/chains?org=National+Highways&q=hard+shoulderRecurrence chains, filtered by organisation name and words. Also limit and offset.

GET /api/v1/chains/c52cded99f756309One chain: its reports, each verified link with both quotes, stated actions and follow-up evidence.