Systems and records
Publishing correction provenance data: keep each fix with its original
By SourceX Editorial · Reviewed by Noah Loul ·
9 sections
Short answer
A media dataset should propagate a published correction by keeping the original version, the corrected version and the correction notice as linked records, never by overwriting the original. A correction manifest records which version revises which, what changed and why, who approved it, whether each version may still be delivered, and the rights attached to each, so a corrected record and its original always travel together.
Provenance trail
Publishing correction
Key takeaways
- Overwriting the original hides the error a correction exists to record, and leaves stale copies elsewhere uncorrected.
- Keep three linked records per correction: the original version, the corrected version and the notice that connects them.
- Not every edit needs a correction record; a written threshold separates corrections, clarifications and updates from silent typo fixes.
- Each version carries its own delivery status and rights state, because a correction can swap a photo, a quote or a source with different terms.
- A manifest passes only when every version link resolves, every correction has an approval, and no withheld version is marked deliverable.
- Export revisions from the CMS early: revision histories can be trimmed or deleted, and the original is the hardest record to recover.
Why does a quietly fixed article teach a model the wrong thing?
When a publisher fixes an error by editing the live article and saving over it, the archive keeps only the corrected text. A dataset built from that archive loses three things at once: the fact that an error was made, what the error was, and the moment it was fixed. Meanwhile the original wording may survive in syndicated copies, search caches and earlier exports, so a model can see the wrong version elsewhere with nothing to say it was withdrawn.
Scholarly publishing settled this question years ago. Crossref's Crossmark documentation describes a correction or retraction as a separate notice with its own DOI that points to the work it updates, and it says in-situ updates that replace the earlier version are not considered best practice because they obscure the record. News ethics points the same way: the SPJ Code of Ethics asks journalists to acknowledge mistakes, correct them promptly and prominently, and explain corrections and clarifications clearly.
For a dataset, the practical rule follows. A correction is an event with two ends, and both ends stay in the archive.
Which changes need a correction record, and which do not?
Write the threshold down before building anything. Crossref limits Crossmark registration to editorially significant changes and lets minor fixes such as spelling or formatting pass without notice, and it recognizes a dozen update types ranging from correction and erratum to retraction and expression of concern. NISO's Recommended Practice RP-45-2024 on retractions, removals and expressions of concern focuses on how metadata for both the original publication and the notice is created, transferred and displayed, which is exactly the problem a dataset inherits.
A newsroom or B2B publisher can use a shorter list. The table below is one workable version; the test for each row is whether a reader relying on the earlier text would have been misled.
Decide who applies the threshold. If reporters classify their own changes, silent fixes tend to absorb real errors. A common pattern is that anyone may make a silent fix, but any change to a figure, name, date, quote or attribution goes to the standards desk, which assigns the change type. The manifest then stores the classifying role as well as the approver, and an audit can sample silent fixes to confirm none should have been corrections.
| Change type | Example | Correction record needed? | Effect on the earlier version |
|---|---|---|---|
| Silent fix | Typo, broken link, formatting | No; keep the revision in the version log only | Earlier version stays deliverable |
| Correction | Wrong figure, name, date or attribution | Yes, with notice text and changed spans | Earlier version deliverable only with the notice linked |
| Clarification | Accurate but misleading wording | Yes, labeled clarification | Earlier version deliverable with the notice linked |
| Update | New facts added to a developing story | Yes, labeled update, not correction | Earlier version remains a valid snapshot |
| Editor's note or retraction | Sourcing failure or story withdrawn | Yes, with approver at senior editor level | Earlier version withheld from delivery unless counsel agrees |
| Legal or rights removal | Court order, settlement, expired photo license | Yes, notice may be internal only | Affected versions or assets withheld |
What belongs in the correction and version provenance manifest?
The manifest is a small table that sits beside the article archive and holds one row per version plus one row per correction. It borrows existing vocabularies where they fit. IPTC's ninjs news format defines a version, firstCreated and versionCreated timestamps, a pubStatus of usable, withheld or canceled, an internal edNote and a rightsInfo object. Schema.org gives CreativeWork a correction property that can point to a CorrectionComment. W3C PROV-O supplies wasRevisionOf for saying that one version is a revised form of another.
At a glance
What belongs in the correction and version provenance manifest
Field names below are suggestions. What matters is that every link resolves and every rule is checked by code rather than by eye.
Store the full text of every published version rather than only a diff. Diffs are compact, but a dataset user needs to read any version on its own, and a diff chain breaks if one link is lost. The changed_spans field gives the precise edit for each correction, so nothing is lost by keeping whole versions, and the archive becomes easy to check: any version can be compared with what readers saw on the day.
| Field | Maps to | What it records | Validation rule |
|---|---|---|---|
| work_id and version_id | ninjs uri and version | One article and one saved published state of it | version_id unique; work_id stable across versions |
| version_created | ninjs versionCreated | When this version went live | Later than the version it revises |
| revision_of | PROV wasRevisionOf | The version this one replaced | Must resolve to an existing version of the same work |
| correction_id and change_type | schema.org correction | The notice and its class from the threshold table | Required when change_type is not silent fix |
| changed_spans | Manifest only | Before and after text with character offsets in each version | Before text found in the earlier version, after text in the later one |
| notice_text and notice_location | CorrectionComment text | Published notice wording and where it appears | Matches the text published with the corrected version |
| approval | ninjs edNote for internal detail | Approving role, time and reason source | Approval time before the corrected version went live |
| delivery_status | ninjs pubStatus | usable, withheld or canceled for dataset use | Withheld or canceled versions excluded from every delivery |
| rights_state | ninjs rightsInfo | License references for text, photos and embeds in this version | Every embedded asset has a license reference valid for the use |
| downstream_copies | Manifest only | Syndication partners, feeds and earlier dataset releases that carry an earlier version | Each copy has a notification status after a correction |
Which approvals and rights conditions change what may be delivered?
A correction is not only a text change. It can replace a photo whose license did not cover the use, remove a quote at a source's request, or follow a legal complaint. Each of those changes the delivery status of a version, not just its wording, so the manifest records the reason and the approving role alongside the edit.
Three rules cover most cases. A version withheld for legal or rights reasons is never delivered, even as the earlier end of a correction pair; the notice row still travels so the gap is explained. A version that embedded an asset licensed only for display on the publisher's own site cannot go into a dataset with that asset intact, even if the text is fine. And an approval recorded after the corrected version went live is a process failure that a reviewer has to resolve, not a field to backfill.
Territory and term limits matter for wire copy and agency photos. If a story carried agency text or images under a license limited by region or time, each version's rights_state should say so, and the asset-level license is what decides whether that version can be offered at all. A rights reviewer signs off on those rows before any delivery.
What does a passing manifest look like? One corrected figure (illustrative)
This example is illustrative. The publication, story and figures are invented to show the structure and do not describe any real archive.
A trade publication runs a story reporting that a regional manufacturer's output rose 18 percent year over year. Two days later the company points out that the figure was 8 percent; the reporter had misread a table. The standards editor approves a correction, the story is republished with the figure fixed and a notice appended at the foot, and the manifest gains a correction row linking the two versions.
| Row | Key values | Link | Check result |
|---|---|---|---|
| Version 1 | version_id v1, delivery_status usable, rights_state staff text and staff photo | None; first version | Pass |
| Version 2 | version_id v2, created two days after v1, same rights_state | revision_of v1 | Pass: later timestamp, link resolves |
| Correction C1 | change_type correction, changed span 18 percent to 8 percent | Connects v1 and v2 | Pass: before text found in v1, after text in v2 |
| Notice | An earlier version of this story misstated the increase in output, which was 8 percent | notice_location foot of v2 | Pass: matches published text |
| Approval | Standards editor role, approved before v2 went live, reason source company email | Attached to C1 | Pass |
| Downstream copies | One syndication feed and one earlier dataset release carried v1 | Listed on C1 | Pass: both marked notified |
Which checks catch a failing manifest?
Change the same story slightly. The CMS kept only the corrected text, so version 1 was rebuilt from a syndication copy that already differs in its headline. The correction row exists but its before text cannot be found in the rebuilt version. The approval is dated the day after the republication, and a photo swapped in version 2 has no license reference. Every field is filled, but the record cannot prove what was published or that the fix was approved.
The checks below run on every work before a release. A work that fails the first three is withheld until the original is recovered from a trustworthy copy; one that fails a rights check is withheld until a rights reviewer clears it.
- Chain check: every revision_of resolves, versions of one work form a single chain, and timestamps increase along it.
- Span check: each correction's before text appears in the earlier version and its after text in the later one.
- Source check: an earlier version rebuilt from a copy outside the CMS is flagged with the copy it came from.
- Threshold check: every non-silent change has a correction row and a change_type from the written list.
- Approval check: approval time precedes the corrected version's go-live time.
- Notice check: published notice text matches notice_text exactly.
- Delivery check: no withheld or canceled version is included in a delivery, and every pair carries its notice row.
- Rights check: every embedded asset in a delivered version has a license reference valid for dataset use.
How do you get every version out of the CMS before it disappears?
The original version is the record most likely to be lost. WordPress, for example, exposes revisions through its REST API at a revisions route under each post, and its reference notes that deleting a revision requires a force flag because revisions do not go to the trash. Some sites also cap how many revisions are kept, and migrations between content systems often carry only the latest text. Export the revision history before any cleanup or platform move, and store published versions separately from drafts and autosaves, which are not published states.
Where the CMS never kept revisions, look for other trustworthy copies: the print edition file, the newsletter send, the syndication feed archive, or an earlier dataset release. Record which one each rebuilt version came from, and treat a rebuilt version as lower confidence than one exported directly.
What happens when a correction lands after a dataset has shipped?
Corrections keep arriving after delivery, so the manifest needs a way to reach copies already in a buyer's hands. The downstream_copies field is that mechanism: when a new correction row is written, every listed release that carried the earlier version gets a supplemental entry with the new version, the notice and any change in delivery status.
If the correction withholds a version for legal or rights reasons, the supplemental entry becomes a withdrawal and follows whatever removal process the license agreement sets. Agree that process before the first delivery, including how quickly the buyer applies it and how it is confirmed, rather than improvising it when a complaint arrives.
Keep the release notes honest about timing as well. A release built on a given date reflects the corrections known on that date, so record the manifest snapshot each release was built from. When a buyer asks why two releases differ for the same story, the answer is then a lookup rather than an investigation.
Bringing a corrected archive to a licensing review
Publishers can license archives through SourceX, which manages rights review, contracting, preparation, delivery and payment for them. For a publisher, a correction manifest is useful evidence in that review because it shows which versions are clean, which carry assets with limited rights, and which must stay out.
Data moves only after the publisher has approved scope and terms and signed; from then on, use and sublicensing of the de-identified dataset are limited to what that agreement permits. A publisher can see whether its archive qualifies through a short assessment, without sending any articles.
Frequently asked questions
What should a correction and version provenance manifest include?
One row per published version with its timestamp, the version it revises, its delivery status and rights state, and one row per correction with change type, changed spans, notice text, approval and the downstream copies that carried the earlier version. Each link and rule should be checked by code.
Should the original, uncorrected version stay in a training dataset?
Usually yes, paired with the correction notice and the corrected version, because the pair records what was wrong and how it was fixed. Versions withheld for legal or rights reasons are the exception and stay out of every delivery.
Which approval, rights condition or revision changes what may be delivered?
A retraction or legal removal withholds the affected version, an embedded photo or wire text licensed only for on-site display blocks that version until the asset is removed or cleared, and a correction approved after go-live is held for review. Silent typo fixes change nothing.
How do we tell a correction from an update to a developing story?
A correction fixes something that was wrong when published; an update adds facts that were not yet known. Label them differently in the manifest, because a model should learn that an updated story was accurate at the time and a corrected one was not.
What valid and invalid examples demonstrate the manifest structure?
A valid manifest has a resolving version chain, a correction whose before and after text appear in the right versions, an approval dated before republication and rights references for every asset. An invalid one has an original rebuilt from an unrelated copy, a late approval or an unlicensed swapped photo.
Sources
- Crossref registers corrections and retractions as separate notices with their own DOI, applies Crossmark to editorially significant changes and says in-situ updates obscure the record. Source
- The SPJ Code of Ethics asks journalists to acknowledge mistakes, correct them promptly and prominently, and explain corrections and clarifications clearly. Source
- NISO RP-45-2024 sets practices for creating, transferring and displaying metadata for both an original publication and its retraction, removal or expression of concern. Source
- Schema.org defines a correction property on CreativeWork and a CorrectionComment class. Source
- IPTC ninjs defines version, firstCreated, versionCreated, pubStatus (usable, withheld, canceled), edNote and rightsInfo for news objects. Source
- W3C PROV-O models a revision as a derivation in which the result is a revised version of an original entity. Source
- The WordPress REST API exposes post revisions under each post, and deleting a revision requires force because revisions do not support trashing. Source
See if your company qualifies
A short company assessment. No data uploads are needed.