Skip to content

AI uses for records

Expert correction knowledge dataset: what 1,545 public fixes omit

By SourceX Editorial · Reviewed by Noah Loul ·

9 sections

Short answer

An expert correction knowledge dataset links four things: the answer that was wrong, the corrected answer, the evidence the expert relied on and the role that approved the change. Public correction records show where that chain breaks. Of 1,545 corrections in the Retraction Watch data, 97% link both versions and 99% state what was wrong, but only 13% name a party that checked the fix.

Agent run

Expert correction knowledge

This page covers expert correction knowledge dataset: what 1,545 public fixes omit.Illustrative

Key takeaways

  • Treat an expert's correction as four linked records, the original answer, the corrected answer, the evidence as it stood at review and the approval, plus a status that can change.
  • In the Retraction Watch data generated on 27 October 2026, 1,492 of 1,545 correction records carry DOIs for both the original paper and the notice.
  • Only 205 of those corrections (13.3%) carry a code naming an investigating journal, institution, agency or third party, against 51.6% of retractions.
  • 756 of the 1,545 tracked corrections concern a paper that was later retracted, so a correction label must be able to change after the fix.
  • Ask a seller to show approver role and authority on every approved correction before paying for volume.

What should an expert correction knowledge dataset preserve?

Four linked records and one status. An expert correction knowledge dataset keeps the answer that was wrong, the answer that replaced it, the evidence the expert relied on as it stood at the moment of review, and the approval: which role signed off, under what authority and on what date. The status then records whether the correction itself was later confirmed, revised or withdrawn.

Each piece serves a different use. The pair of answers trains revision and preference behavior. The evidence lets a grader check that the new answer is supported rather than merely newer. The approval tells a buyer whether the fix came from someone entitled to make it, which matters when one knowledge base mixes subject experts, support agents and automated edits. Without the status, a model can learn from a fix that was later reversed.

This page measures how often a mature public correction system records those links, then turns the gaps into an acceptance test a buyer can attach to a request. The field-level annotation scheme for one reviewer's fix has its own guide, as does manifest design for corrected media archives; neither counts how often real correction records carry each link, which is the question here.

Why published research corrections make a fair test

Scholarly publishing runs one of the most formal public correction systems, so its gaps set a floor for what to expect in less formal settings. The original article and the correction notice usually each carry a DOI, journals work from published editorial guidance, and an outside database codes why each notice it tracks was issued.

That guidance asks mainly for links and dates. The ICMJE recommendations on corrections and version control say a journal should publish a correction notice that details the changes and cites the original publication, and post a new version of the article with details of the changes and the dates they were made. COPE's retraction guidelines ask more of retractions: the notice should state who is retracting the article and the reasons.

The Retraction Watch database, acquired by Crossref in September 2023 and published as an open CSV, codes each notice with reasons from a controlled vocabulary. Its README says the data is updated every working day and that corrections and expressions of concern are included but are not as comprehensive as retractions. That limit shapes everything below: the corrections in the file are the ones Retraction Watch chose to track, not every correction journals have issued.

What 1,545 correction records show, field by field

They link versions reliably and say what went wrong, but rarely say who checked the fix. The snapshot generated on 27 October 2026 holds 73,399 notices: 67,958 retractions, 3,733 expressions of concern, 1,545 corrections and 163 reinstatements. The corrections were issued between 1981 and 2026.

Of the 1,545 corrections, 1,492 (96.6%) carry a DOI for both the original paper and the notice, and 1,528 (98.9%) carry at least one reason code that names a cause, such as an error in an image, in the text or in the data. Just 36 notices (2.3%) sit behind a paywall.

The reviewer is the weak link. Only 205 corrections (13.3%) carry a code saying that a journal or publisher, the authors' institution, a third party or the U.S. Office of Research Integrity investigated. Retraction notices name an investigating party in 51.6% of cases and expressions of concern in 61.8%, so the lightest-weight fix is also the least attributed one. Put side by side, 51.6% of retractions carry an 'Investigation by' code; the authors' own institution accounts for 4.8%. Requiring both DOIs, a cause, a named investigating party and an open notice leaves 191 corrections, or 12.4%.

The gap has not narrowed with time. Corrections issued from 2015 to 2019 name an investigating party in 17.3% of cases, those from 2020 to 2022 in 6.8%, and those from 2023 onward in 15.2%. The three notice types compare as follows:

What 1,545 correction records show, field by field
Field in the recordCorrections (1,545)Retractions (67,958)Expressions of concern (3,733)
DOI for both the original and the notice96.6%89.4%97.8%
At least one cause code98.9%89.1%96.9%
Names an investigating party13.3%51.6%61.8%
Notice behind a paywall2.3%2.2%3.0%
Both DOIs, a cause, an investigating party and an open notice12.4%49.5%57.3%
Median days from paper to notice6554941,054
A dot plot compares corrections, retractions and expressions of concern in the Retraction Watch data on four fields: both versions linked, a cause recorded, an investigating party named, and all four with an open notice; corrections sit lowest on naming an investigating party, at 13.3%.
Share of notices carrying each field, Retraction Watch data generated 27 October 2026 (Crossref, commit 2c24a2f). Codes are assigned by Retraction Watch.

A correction is often not the last word

In this data, the paper behind a correction was later retracted almost half the time. Of the 1,545 corrections, 756 (48.9%) share their original-paper DOI with a retraction record, and 557 carry the reason code Updated to Retraction. Because Retraction Watch tracks corrections selectively, this share describes the corrections it chose to follow, which lean toward troubled papers; it is not a rate for corrections in general.

Two lessons carry over to company knowledge. A correction needs a status of its own that can change after the fix, with a link to whatever superseded it. And the fix can arrive long after the original was in use: the median correction here appeared 655 days after the paper, and 965 of the 1,545 (62.5%) appeared more than a year later. Anything a model learned from the original in that window was the uncorrected version.

Corrections later followed by a retraction named an investigating party less often than the rest: 53 of 756 (7.0%) against 152 of 789 (19.3%). The data cannot say why. It does show that the fixes which turned out not to be final were the ones least likely to record who had examined them.

The correction evidence map: seven fields and how the public record fares

Use the map as the field list for any correction dataset and as the order in which to check a seller's sample. Each row pairs a field with its nearest public counterpart, the share of the 1,545 corrections that carry it, and the check to run before buying volume. Approver authority comes first because it is the field the public record most often lacks.

Two rows have no clean public measure. The notice text is the only evidence most journal corrections expose, and the file cannot show whether a notice attaches the data or images behind the fix. Reason codes are assigned by Retraction Watch from each notice, so the 98.9% shows that a cause could be read from the notice, not that the journal filled in a structured field. The seven fields and their checks:

The correction evidence map: seven fields and how the public record fares
FieldWhat it should holdPublic counterpartShare of 1,545 correctionsBuyer check on a sample
Approver and authorityRole, pseudonymous ID and the policy that lets that role approve this kind of fixA code naming the investigating party13.3%Every approved correction names a role entitled to approve it
Original answerID and exact version of the answer that was wrongOriginal-paper DOI97.9%Each link opens the version that was corrected, not the current one
Corrected answerID and version of the replacement, with what changedNotice DOI97.2%A diff or changed-field list comes with every pair
What was wrongError type from a fixed list plus a one-line rationaleReason codes assigned by Retraction Watch98.9%No record relies on a free-text label alone
Evidence as reviewedPointers to the sources the expert used, frozen at review timeThe notice text; underlying data is rarely attachedNot measuredA second reviewer can open every cited source
AccessWhether the buyer may read the notice and the evidenceNotice not behind a paywall97.7%Restricted evidence is listed, not silently dropped
Later statusConfirmed, revised or withdrawn, with a link to what replaced the fixA later retraction of the same paper48.9% later retractedEvery correction carries a dated status

The evidence-grounding test: could a second expert reach the same fix?

A complete record should let them: a correction record is complete when a second qualified reviewer, given only the record, reaches the same corrected answer or can point to exactly where they disagree. That is a stronger test than counting filled fields, because a record can fill every field and still cite evidence that does not support the change.

Run it on a stratified sample before buying at scale. Draw corrections across error types and approver roles, hide the corrected answer, and give a second reviewer the original answer, the evidence pointers and the rule the first reviewer cited. Record one of four outcomes: same fix, same fix on different evidence, a different fix, or not decidable from the record.

Keep the disagreements in the dataset with both readings instead of resolving them quietly. A contested correction is a legitimate label, and it is the case a model most needs to see. Set thresholds before the sample arrives; the figures below are illustrative starting points, not industry standards:

  • Same fix reached from the record alone: at least 8 in 10 sampled corrections (illustrative).
  • Not decidable from the record: no more than 1 in 10, each traced to the missing field (illustrative).
  • Approver role present and entitled to approve: every approved correction, without exception.
  • Evidence pointers that still open to the version reviewed: at least 9 in 10 (illustrative).
  • Corrections later revised or withdrawn: all listed with dates, whatever their share.

The strongest objection: the corrected answer is the signal, the paperwork is overhead

For cosmetic edits the objection holds. A typo fix or a repaired link needs the before-and-after pair and little else, and demanding approver authority for every spelling change would make a dataset expensive without making it better.

It fails where the correction changes a fact, a figure or an instruction someone will act on. Relevant text can be outdated, unauthorized or too thin to support a confident answer, and a model trained on the corrected version inherits whichever of those problems the fix carried. The research record shows both failure modes at scale: almost half the tracked corrections were followed by a retraction of the same paper, and 86.7% do not name an investigating party. A buyer who cannot tell an approved correction from an unreviewed edit is training on assertions.

A workable rule scales the record to the consequence: full evidence and authority for corrections to facts, amounts, policies and safety-relevant steps; the answer pair and a reason code for cosmetic edits; and a flag that separates the two so neither class contaminates the other.

Where company archives keep correction evidence, and how to request it

Inside companies, expert corrections sit in systems that already record versions and actors: knowledge base article histories, edits to support macros and templates, quality review tools where a senior reviewer overrides an agent's answer, reopened tickets with a supervisor's note, and approval steps in document systems. What usually goes missing is the link from the fix to its evidence, and the approver's authority, which lives in policy documents or org charts rather than in the record.

Attribution should travel as a role and a pseudonymous ID, not a name. A buyer needs to know that a reviewer with authority over refund policy made the change, not who that person is, and a stable pseudonym is enough to measure agreement between reviewers and drift over time. Names, emails and comments that identify staff or customers should be removed or replaced before any sample leaves the company.

SourceX takes its licenses straight from the companies that hold the records and licenses the resulting datasets to AI developers in its own name, so correction data is sourced for a request rather than kept in stock. Buyers describe what they need on the request form, covering scope, volume, language, format, timing and licensing terms, and SourceX follows up about potential sources. A request for correction data is easiest to answer when it attaches the field map above as the acceptance test, states the approver-authority rule, and says what share of contested corrections should be kept.

Methods note: the Retraction Watch snapshot, counting rules and limits

Data: retraction_watch.csv from Crossref's retraction-watch-data repository on GitLab at commit 2c24a2f (data generated 27 October 2026), cloned on 28 October 2026 and read by a script (rw_corrections.py) that loads only the record ID, notice and paper dates, the two DOIs, the notice PubMed ID, the notice type, the reasons and the paywall field. Titles, author names, institutions, countries, URLs and notes were never read into the analysis; the saved file holds record IDs, yes-or-no flags, notice years and day gaps.

Rules: the population is every row whose notice type is exactly Correction, compared with Retraction and Expression of concern rows. A DOI counts as present when the trimmed, lowercased value starts with 10. Reasons are split on semicolons, trimmed, with a leading plus sign removed. A record names an investigating party when any reason begins with Investigation by. The companion article "External internal claim provenance dataset: what retractions record" counts five investigation codes in the same release and reports 4.8% of retractions citing the authors' own institution and 51.6% carrying at least one of the five. Fourteen codes count as procedural rather than causes: Updated to Retraction, Updated to Correction, Updated to Expression of Concern, Upgrade/Update of Prior Notice(s), Date of Article and/or Notice Unknown, the three Notice codes (Limited or No Information, Lack of, and Unable to Access via current resources), Removed, Temporary Removal, Retract and Replace, No Further Action, EOC Lifted and Doing the Right Thing; every other code that is not an investigation code counts as a cause. A paywalled notice has Paywalled equal to Yes. A later retraction means the correction's original-paper DOI equals the original-paper DOI of a retraction row. Day gaps are exact date differences; rows with a missing or unparseable date or a negative gap are left out of the day figures, which excluded one retraction row. Paywall values of Unknown (221 retraction and 4 expression-of-concern rows, no corrections) count as not paywalled. Medians ending in .5 round half up, and percentages are rounded once from exact counts.

Limits: reason codes are Retraction Watch's reading of each notice, not fields the journal completed, and none of the codes in this snapshot records an editor approving an authors' fix, so 13.3% measures how often a reviewing party beyond the routine editorial process is recorded, not how often anyone checked. Corrections are tracked selectively, so shares describe this set rather than all journal corrections. Not measured: whether notices attach underlying data or images, the quality of correction text, the delay between a problem being raised and the notice, and company archives, for which SourceX has published no measurements.

Frequently asked questions

What is an expert correction knowledge dataset?

A set of corrections in which each record links the original answer, the corrected answer, the evidence the expert relied on and the approval, plus a status that can change later. It supports training on revisions, grading answers against evidence and checking whether a fix came from someone entitled to make it.

How often do published corrections say who checked the fix?

Rarely. In the Retraction Watch data generated on 27 October 2026, 205 of 1,545 correction records (13.3%) carry a code naming an investigating journal, institution, agency or third party, against 51.6% of retraction records. Retraction Watch assigns those codes after reading each notice.

Should a correction dataset include the reviewer's name?

No. A role, a pseudonymous ID and the policy that gives that role authority are enough to judge a correction and to measure agreement between reviewers. Names, emails and identifying comments should be removed or replaced before a sample is shared.

What happens when a correction is itself withdrawn?

Keep both records and give the correction a status with a date and a link to what replaced it. In the Retraction Watch data, 756 of 1,545 tracked corrections concern a paper later retracted, so a fixed corrected label can mislead a model.

How is a correction dataset different from an edit history?

An edit history shows that text changed and who typed the change. A correction dataset adds why it changed, the evidence that supports the new version and whether the editor had authority to make it, which is what a grader or a buyer needs before trusting the fix.

Sources

  • Retraction Watch data (Crossref) at commit 2c24a2f, generated 27 October 2026: 73,399 notices including 1,545 corrections; the source of every count of DOIs, reason codes, paywalls, later retractions and day gaps in this article. Source
  • The repository README says Crossref acquired the Retraction Watch database in September 2023, that it is updated every working day, and that corrections and expressions of concern are included but are not as comprehensive as retractions. Source
  • Retraction Watch's change log describes the investigation reason codes: investigation by the journal or publisher, by the company or institution listed as an author's affiliation, by a third party other than ORI, and by the U.S. Office of Research Integrity. Source
  • ICMJE recommends that a journal publish a correction notice detailing the changes and citing the original publication, and post a new article version with details of the changes and the dates they were made. Source
  • COPE's retraction guidelines say notices of retraction should state who is retracting the article and the reasons for retraction. Source
  • SourceX's request page says it licenses business data directly from companies and licenses it to buyers in its own name, and asks buyers to share scope, volume, language, format, timing and licensing needs. Source

Tell us what data you need

Share scope, volume, format, timing and licensing needs. Datasets are sourced on request.

See if your company qualifies