Systems and records
External internal claim provenance dataset: what retractions record
By SourceX Editorial · Reviewed by Noah Loul ·
10 sections
Short answer
An external internal claim provenance dataset pairs each public assertion with the records that bear on it, a dated status history and the party that established each change. Tested on 73,399 Retraction Watch records, the fields matter: only 4.8 percent of retractions cite the asserter's own investigation, the median took 541 days, and 52.1 percent of expressions of concern remain unresolved.
Provenance trail
External internal claim
Key takeaways
- Attribute both sides: record who asserted a claim, where and when, and which internal record supports or contradicts it.
- Record who established each status change; only 4.8 percent of 67,958 retractions cite the asserting institution's own investigation, against 51.6 percent with some investigation code.
- Index status by date: the median retraction came 541 days after publication, so snapshots hold claims later withdrawn.
- Keep open and reversed statuses: 52.1 percent of expressions of concern remain unresolved and 163 notices were reinstated.
- Allow reason not recorded as a value; 17.6 percent of retractions carry notices with limited or no information.
What separates an external claim from an internal fact in a knowledge dataset?
Attribution and status: who asserted the statement, where and when, which record the organization holds on the same point, and what the statement's verification status was on a given date, together with who established that status. An external claim is anything said for an outside audience, such as a press release, a product page, a sales deck or a vendor's documentation; an internal fact is what the organization's own system of record shows.
Neither side is automatically right. A product page can promise an uptime figure that the incident log contradicts, and a ledger entry can be the error that a public correction later fixes. A knowledge assistant that answers from both without attribution blends them into one confident sentence, and the reader cannot tell a marketing line from a measured value.
The dataset that prevents this pairs each assertion with the records that bear on it and a dated status history. This page tests what such a history must contain on the largest public corpus of statements that were later corrected or withdrawn: the Retraction Watch database, 73,399 records in the 2026-10-27 release.
Why test claim provenance on retraction notices?
Because every record pairs an original public statement with a later change in its status, dated, typed and coded by reason, which is the structure a company's claim ledger needs. The data repository's README says Crossref acquired the database in September 2023 and made it publicly available, that it is updated every working day, and that expressions of concern and corrections are included but are not as comprehensive as retractions.
The release read for this page holds 67,958 retractions, 3,733 expressions of concern, 1,545 corrections and 163 reinstatements, each linked to the original publication's date and, usually, its DOI. Reasons come from a controlled vocabulary, and the database's own field guide warns that some reasons describe the nature of the notice rather than the actual reason, citing a label used when an article was removed with no notice at all.
The analogy has limits: a journal article is not a press release, and an editor is not a compliance officer. Companies keep the same record types in their own formal disclosures, though. Under Item 4.02 of Form 8-K, a public company reports when its board, a board committee or authorized officers conclude that previously issued financial statements should no longer be relied upon, and separately when its independent accountant advises that action is needed to prevent reliance on an audit report (paraphrase). One filing records both the reversal and who prompted it.
Who established that the public statement was wrong?
Usually someone other than the organization that made the statement, and often nobody the notice names. Of the 67,958 retractions, 30,878 (45.4 percent) cite an investigation by the journal or publisher, 17,946 (26.4 percent) one by a third party, 1,541 (2.3 percent) official misconduct findings, and only 3,277 (4.8 percent) an investigation by the authors' own company or institution. None of the five investigation codes the script counts appears on 32,882 (48.4 percent). Side by side, 4.8% of retractions carry the code for an investigation by the authors' own institution; 51.6% carry some investigation code.
Self-correction is rarer still when it stands alone: 1,354 retractions name the authors' company or institution as the only investigating party among those five codes. Papers with a company affiliation look much the same. Among 1,703 retractions where an affiliation contains a corporate designator such as Inc, GmbH or Ltd, 4.4 percent cite the institution's own investigation and 44.6 percent cite no investigation at all.
Authority to change the status also sits outside the asserting organization. COPE's position is that institutional findings can inform an editor's retraction decision but do not determine it (paraphrase). For a company dataset, that makes the establishing party a required field with fixed values, because corrected on its own conflates a company catching its own error with a regulator, auditor or customer exposing it.
How long does a public claim stay live before its status changes?
Long enough that any snapshot of public statements contains claims that will later be withdrawn. Across 61,048 retractions with usable dates, the median time from original publication to retraction was 541 days; 34.9 percent came within a year, 26.3 percent after more than three years and 4.3 percent after more than ten.
Status therefore has to be indexed by date. A question about what a company said in March needs the statement as it stood in March, the status in force then, and the date the organization learned otherwise, which can differ from the date it said so publicly. A dataset that stores only the current status teaches a model to answer historical questions with hindsight.
The lag count leaves out 6,909 retractions coded as having an unknown article or notice date and one whose retraction date falls before its publication date; both exclusions are printed by the script.
What happens when a status is disputed or never resolved?
Many stay open for years. Of 3,704 expressions of concern with a DOI and a date, 1,930 (52.1 percent) have no later retraction, correction or reinstatement for the same DOI; the median open concern was 1,737 days old on 2026-10-27, and 716 were more than five years old. COPE guidance describes an expression of concern as a notice that can be interim, later followed by a correction or retraction, or can be the final outcome for an article that will not be updated (paraphrase).
Status also moves in both directions, and the asserting party sometimes resists. Authors objected in 4,011 retractions (5.9 percent) and did not respond in 3,497 (5.1 percent); 163 reinstatements reversed an earlier notice, and 2,145 records note an earlier notice upgraded to a retraction. A model that treats any concern as proof of falsehood will be wrong on the reversals.
Some status changes arrive without a reason. Of the retractions, 11,986 (17.6 percent) are coded as notices with limited or no information, and 10,471 (15.4 percent) carry that code with no investigation code either: the status is known, but the reason and the establishing party are not. A claim ledger needs reason not recorded as a legitimate value, so that nothing downstream invents one.
The claim-status ledger: seven fields that keep assertion and record apart
Seven fields keep a public assertion and the records that bear on it separate, each tied to what the Retraction Watch counts show and to where a company would find the equivalent. The last column is the behavior a trained or evaluated model should show.
The ledger, with evidence from the 2026-10-27 release:
| Field | What it separates | Retraction Watch evidence | Enterprise source | Model behavior |
|---|---|---|---|---|
| Assertion: text, channel, date, asserting party | The public statement from what records show | Every record ties a notice to an original publication date and usually a DOI | Press releases, product pages, sales decks, vendor documents | Quote the claim with its source and date, never as plain fact |
| Status event: type and date | Asserted, under concern, corrected, withdrawn, reinstated | 67,958 retractions, 3,733 concerns, 1,545 corrections, 163 reinstatements | Correction notices, restatements, recalls, errata, status-page updates | Give the status in force on the question's date |
| Establishing party | Self-correction from outside exposure | 4.8 percent cite the asserter's institution; 48.4 percent name none of five investigation types | Internal audit, QA, customer complaint, regulator, auditor | Name who established the status, or say it is unrecorded |
| Grounding record | Which internal record supports or contradicts the claim | Absent: notices give reason codes, not record identifiers | Ledger entry, incident ticket, test report, contract clause | Cite the record identifier and version behind an internal fact |
| Reason disclosed | A known reason from an undisclosed one | 17.6 percent of retractions coded limited or no information | Change logs, approval notes, correction memos | Say the reason is not recorded rather than inventing one |
| Dispute and reversal | Settled from open or contested | 52.1 percent of concerns still open; 5.9 percent author objections | Appeals, disputes, pending reviews | Report open or disputed status without picking a side |
| Valid-from and learned-on dates | When it held from when it was known | Median 541 days from publication to retraction | Posting dates, period-end dates, discovery dates | Apply as-of logic to dated questions |
How would a buyer train and grade a model on claim-status records?
As cases whose gold answer is an attributed, dated statement rather than a bare fact. An evaluation item gives the model a corpus holding an external claim and an internal record, plus a question with an as-of date; the gold answer names who asserted what, through which channel and when, what the record shows, the status on that date and who established it, and it says so plainly when the reason was never recorded.
Scoring follows the ledger: credit for correct attribution, correct status as of the question's date, a cited grounding record for any internal fact, and an explicit flag for open or disputed items; a penalty for any invented reason. For training, the same pairs carry labels such as consistent, contradicted, unverifiable and superseded, and the set should include reinstated claims so the model does not learn that every concern means false.
A buyer can test a supplier's sample in the same terms: the share of status events with an establishing party, the share with a grounding record identifier, whether originals survive next to their corrections, and whether dates distinguish valid-from from learned-on. Keeping a corrected record linked to its original is covered in the guide to publishing correction provenance, and versioned numbers in the guide to financial restatement datasets; this page covers attributing a claim and its status in a synthesized answer.
What could make a claim-status dataset misleading?
Five weaknesses can mislead a buyer, and the strongest is treating a missing correction as verification.
- Absence is not verification. The README says corrections and expressions of concern are covered less completely than retractions, so a claim with no notice is unreviewed, not confirmed.
- Codes describe notices, not causes. Labels such as limited or no information describe the notice itself, so the reason field records what was said publicly, which can differ from what an investigation found.
- Internal records can be the error. A ledger or system of record is evidence, not ground truth; the ledger attributes status and never ranks internal over external by default.
- Internal records differ in authority. A finance owner's posted figure and a sales representative's note are both internal; record which role may confirm each fact.
- Rights travel with the records. Internal documents that contradict public statements may be confidential, privileged or personal; each needs rights and privacy review before licensing, and counsel should review the specific records.
Asking a supplier for claim-status records
Name the claim types and channels, the status values, the establishing-party values, the grounding-record links and the dates you need, and attach the acceptance test you will run on a sample. SourceX looks for companies holding matching records after a buyer describes the need, since it keeps no stock of datasets, and it licenses what it sources to the buyer in its own name.
Candidate records include corrected press releases linked to their internal approvals, product specification change logs, restatement workpapers, support macros revised after a wrong answer, and status-page updates tied to incident tickets. Whether a particular company's records carry these links, and whether they can be licensed, is settled case by case; nothing here implies that such data is available today.
Methods note: file, counting rules and what was not measured
Data: retraction_watch.csv from gitlab.com/crossref/retraction-watch-data at commit 2c24a2f (the release generated on 2026-10-27), read with git show (SHA-256 ec6d6d5d8aee2c87e0582849943e3cf99f211d5c157927ad0f50a8445d42ac07); one row is one record. The script, retraction_status_census.py, uses the Python standard library with no randomness; the saved snapshot keeps record ID, nature, dates, original DOI, the company flag and reason codes, and drops titles, author names and every other field.
Rules as applied: Reason is split on semicolons and a code counts only as an exact whole item. The five investigation codes are Investigation by Company/Institution, Investigation by Journal/Publisher, Investigation by Third Party, Investigation by ORI and Misconduct - Official Investigation(s) and/or Finding(s). Dates are read as month/day/year and lags are whole days; a retraction leaves the lag count when a date is missing or unparseable, when it carries the code Date of Article and/or Notice Unknown, or when its lag is negative. An expression of concern is open when no retraction, correction or reinstatement with the same lowercased DOI (starting 10.) is dated on or after it. A record is company-affiliated when any semicolon-separated affiliation contains, as a whole case-sensitive word, Inc, Incorporated, Corp, Corporation, LLC, GmbH, AG, plc, PLC, Ltd or Limited. Shares are exact fractions rounded once, half up; an even-count median is the mean of the two middle values, rounded half up to a whole day. The companion article "Expert correction knowledge dataset: what 1,545 public fixes omit" reads the same release with a broader rule, any reason beginning Investigation by, and also finds an investigating party on 51.6 percent of retractions.
Not measured: whether any retraction was justified, any institution's internal records, author identities, and how far academic notices predict corporate behavior. The company-affiliation rule is a text match, so it includes some non-company bodies and misses companies without a designator. The ledger and the evaluation design are editorial frameworks drawn up for this page on 2026-10-27.
Frequently asked questions
What is a claim provenance dataset?
A dataset that stores each statement with who made it, where and when, the records that support or contradict it, and a dated history of its status. Its gold answers are attributed statements rather than bare facts. It teaches a model to say who claims what and how far that has been verified.
How is an external claim different from an internal fact?
An external claim is made for an outside audience, such as a press release or product page. An internal fact is what the organization's own records show, such as a ledger entry or incident ticket. Either can be wrong, so a dataset should attribute both instead of ranking one above the other.
Why does verification status need a date?
Because status changes long after a statement is made. In the Retraction Watch release of 2026-10-27, the median retraction came 541 days after publication, and 26.3 percent came after more than three years. A question about a past date needs the status in force then.
Who should a dataset record as establishing a correction?
The party whose finding changed the status: the organization's own review, a publisher or platform, a third party, an auditor or a regulator, or unrecorded. In the retraction data, only 4.8 percent of retractions cite an investigation by the authors' own institution, while 51.6 percent carry some investigation code. Without this field, self-correction and outside exposure look identical.
Does the absence of a correction mean a claim is verified?
No. Correction databases are incomplete; the Retraction Watch README says corrections and expressions of concern are covered less fully than retractions. A claim with no notice should be labeled unreviewed unless a grounding record confirms it.
Sources
- The Retraction Watch data repository, distributed by Crossref, holds the database Crossref acquired in September 2023, updated every working day, with expressions of concern and corrections covered less comprehensively than retractions; its CSV is the file counted here. Source
- Retraction Watch's field guide says some reasons describe the nature of the notice rather than the actual reason, such as a label for an article removed without any notice. Source
- COPE guidance describes an expression of concern as a notice that can be interim, followed by a correction or retraction, or the final outcome (paraphrase). Source
- COPE's position is that institutional findings can inform but do not determine an editor's retraction decision (paraphrase). Source
- Form 8-K Item 4.02 requires disclosure when a company concludes that previously issued financial statements should no longer be relied upon, and when its independent accountant advises action to prevent reliance on an audit report (paraphrase). Source
- SourceX sources datasets on request rather than holding them in stock and licenses them to buyers in its own name. Source
Tell us what data you need
Share scope, volume, format, timing and licensing needs. Datasets are sourced on request.