Skip to main content
Defiance International

How study summaries are made

A machine drafts each summary from a study’s published abstract. A second machine compares every sentence with that abstract. No person reviewed any summary before it was published.

Who writes a summary

A machine drafts the summary. A second machine compares each drafted sentence with the abstract and judges whether the abstract supports it.

No person reviewed the summary. Nobody at this business reads each study, and only the abstract is held, never the full paper.

Each record names the drafting model and the checking model, says whether the two come from different providers, and prints the day of the last machine check.

What a summary is made from

The only source is the abstract as its publisher issued it, retrieved from the index named on the record. The record prints the identifier, the retrieval day and the first 12 characters of the abstract’s fingerprint.

A sentence marked as exact is the abstract’s own sentence. Every other sentence is ours, and the study and its abstract belong to their authors and publisher.

What a machine check does not show

A machine check compares the summary with the abstract. It does not test the study, repeat it, or weigh its findings. A summary that matches its abstract can still describe a study that is wrong.

A machine check can itself be wrong. That is why every record carries a report control and every withdrawal is logged in public.

How to read a source check record

A study record with a published summary carries a source check row. Opening it shows the record of the machine comparison, row by row. A row prints only when its stored value exists for that exact version of the record. A missing value removes the row, and nothing is filled in by default.

The source status comes from a scheduled look at the index for a retraction, a correction, a notice of concern or a changed abstract. A look up to 35 days old, inclusive, keeps the record current. Past 35 days the row reads overdue and the summary stays on the page. Past 90 days the summary is withheld until a new look is stored. An internal warning is raised at 32 days.

Every date on a record is the calendar day in Coordinated Universal Time (UTC). The state of a record is worked out from its stored times and the server clock in UTC each time the page is requested, so a record can pass from current to overdue with no edit to the page.

A fingerprint is a hash. The summary fingerprint covers every machine-written string served on the record: the answer line, the plain-language sentences PL1 to PL4 and the technical sections. Text this platform writes around the summary, such as the labels, the boundary line and this page, is outside the fingerprint. The abstract fingerprint covers the abstract as retrieved. A record prints the first 12 characters of each fingerprint, and it receives a new summary version when the covered text changes.

These checks are run by the business that publishes this page and sells research materials. No independent party takes part in them.

Opening sentence
One sentence that stays true when quoted alone: who published the summary, which abstract it was drafted from, when the machine comparison ran, and the limits of that comparison.
Source
The abstract the summary was drafted from: its identifier, the index it was retrieved from, the retrieval day and the abstract fingerprint. No other part of the paper is held.
Summary
The summary version and the summary fingerprint. The sentences are ours, not the authors’, except where a sentence is marked as exact.
Passages
How many summary sentences point to a marked passage of the abstract.
Figures
How many figures the summary states. Each one must appear in its marked passage exactly as printed.
Second check
The verdict of a second model on each sentence against its passage, how many sentences were replaced with the abstract’s own sentence, and whether the two models come from different providers.
Re-audit
Printed only when the record was drawn into a sample and a third model judged it again. A record outside the sample carries no such row.
Source status
The day of the last look at the index, what that look found, and the day after which the look goes overdue.
People
No person reviewed the summary. This row reads the same on every record.
Error rate
The error rate of published summaries has not been measured. The row says so until a measured figure exists.
What this does not mean
The limits of the comparison. It does not show that the findings are correct, repeatable or important. It is not an endorsement by the authors, the journal, the publisher or the index. It says nothing about any product or any use in a person. A passed machine comparison can still be wrong.
Pins and links
The policy version, the day of the machine comparison, corrections on this record, and links to this page, the corrections log, the report control and the objection procedure.

Reports and corrections

Any reader can report a sentence. The form takes the sentence and a note. It asks for no name and no address, and stores neither.

A report starts one automatic re-check of the record by machine. While a report is in flight the record stays as published, until that job cycle ends. If the re-check fails, the summary is withdrawn in the same cycle and one line is added to the corrections log.

A log line holds the study record id, the sentence id, the cause, the time and the policy version. Lines are never edited and never removed.

For publishers and authors

A publisher or an author of a study listed here can object to its summary, to a quoted passage or to the listing, or ask for removal. Write to [email protected]. Include the PMID or DOI of the study; the address of the page; what is objected to: the summary, a quoted passage or the listing; and the writer’s connection to the study, as its publisher or one of its authors. The outcome of each objection is entered in the corrections log.

Defiance International, a supplier of chemical sequences for research use. Read it before you repeat it: https://defianceinternational.io/studies/method