Corrections and Removal

Edition label: 2026-10

Corrections policy

Applies to Parallax Publications and the Parallax Research Atlas.

1. The commitment

We correct errors in public, on the record, and without quietly rewriting what was already published. A reader who cites an edition must be able to see, later, exactly what changed in it and why.

2. What counts as a correction

• A factual error — a date, figure, name, sequence, or attribution that is wrong.
• A citation error — a source that does not say what we said it says, or does not exist.
• A quotation error — text presented as a quotation that does not match the source.
• A computational error — a count, percentage, or derived statistic that does not reproduce.
• A provenance error — a document, Bates number, or hash that does not resolve to what we claimed.
• In the Parallax Research Atlas: a schema violation, a broken stable ID, or a manifest checksum that does not match its object.

Changes of interpretation, added context, and revised analysis are not corrections. They are new work and belong in the next edition, clearly dated. Calling a changed opinion a “correction” devalues the word for the cases that need it.

3. How to report one

Write to corrections@disclosureworks.ai. Include the edition ID, the page or object ID, what is wrong, and — if you have it — the source that shows what is right. Anonymous reports are accepted and handled the same way.

You do not need to be the subject of the material to report an error in it.

3a. Feedback that is not a correction

General feedback, such as a comment, a question or a suggestion for future work, also goes to corrections@disclosureworks.ai, with the subject line Feedback. We read it, but we do not log it as a correction, and we do not promise a reply. If you think something we published is wrong, report it as a correction under § 3, so that it gets a severity and a deadline.

4. Triage

We acknowledge within three working days. Severity sets the deadline for fixing:

A severity-1 report that we cannot resolve inside 72 hours gets an interim public notice saying the claim is under review, rather than silence while we work.

5. How a correction is published

Editions are immutable under this publication policy. We do not edit a published edition in place. Instead:

1. A correction notice is written: what was published, what is correct, what evidence changed our mind, and the date.
2. A patch release of the affected edition is published carrying the fix — for example 2026-10-patch-1. The original edition remains reachable at its own URL.
3. The release-history page and the latest pointer are updated.
4. In the Parallax Research Atlas, the manifest, object checksums, and any dependent derived index are regenerated, and the affected stable IDs are named in the notice.

The correction notice is permanent. It is not removed in later editions, and it is not rolled up or summarised away once the fix is old.

6. What we will not do

• Silently amend a published page.
• Withdraw an edition to avoid publishing a notice.
• Correct by deletion where a correction notice is what the record requires.
• Treat a demand to retract accurate, sourced material as a correction request. That is a removal request and is handled under the removals policy.

7. Limits worth stating

Our editions are published under CC BY 4.0 and are designed to be copied, mirrored, and ingested by research agents. A correction propagates to our surface, not to every copy of it. We can tell you what we got wrong; we cannot reach into a mirror, an archive snapshot, or a downstream dataset and fix it there. This is a consequence of the licence we chose deliberately, and it is a reason to get things right the first time rather than to rely on the correction route.

Removals policy

Applies to Parallax Publications and the Parallax Research Atlas.

1. Preference order

Where a problem can be fixed without removal, we fix it without removal:

Correct → redact → restrict → remove.

A correction fixes an error. A redaction removes specific sensitive content while leaving the record intact. Removal takes material out of the public edition entirely, and is the last option because it destroys the citability that the rest of this project exists to create.

That ordering is a default, not a delay tactic. Where the material is a victim identification or a minor’s identity, we act immediately and sort out the tier afterwards.

2. Grounds we act on

We will remove or redact material on any of these grounds:

1. Identification of a victim or survivor of sexual abuse, whether by name, image, or detail sufficient to identify them — including where the identification is present in the underlying public record. That a court file names someone is not a reason for us to amplify it.
2. Identification of a person who was a minor at the time of the events described.
3. Material published in error — anything restricted, sealed, embargoed, or unreviewed that reached a public surface. This is also an incident; see SECURITY.md.
4. Sensitive personal data — government identity numbers, financial account numbers, medical information, home addresses, personal phone numbers and email addresses, and dates of birth of private individuals.
5. A court order requiring removal, or material we learn is under seal.
6. A credible, specific safety risk to an identifiable person created or materially increased by our publication.

Requests under grounds 1, 2, 4, and 6 are honoured for private individuals without requiring the requester to justify themselves further than establishing that the ground applies.

3. Grounds we do not act on

We will not remove material because:

• It is unflattering, embarrassing, or commercially inconvenient to a public figure acting in a public capacity.
• It accurately reports a public record, a court filing, sworn testimony, or an official document, and the requester disputes the underlying record rather than our rendering of it.
• The requester asserts an error without identifying one. That is a correction request, and we would rather have it as one.
• A third party objects to material about someone else, absent grounds 1, 2, or 6.

Where we decline, we say so, we say why, and we tell the requester that they may put a response on the record. We would rather publish a disagreement than quietly bury a request.

4. Process

1. Request to corrections@disclosureworks.ai, identifying the edition, the page or object ID, the material, and the ground relied on.
2. Acknowledgement within three working days. For grounds 1, 2, and 6 we acknowledge and act to suppress the specific material immediately, before completing the review.
3. Review by the publication owner against § 2 and § 3.
4. Decision, in writing, to the requester: granted in full, granted in part, or declined with reasons.
5. Action — the removal or redaction ships as a patch release of the affected edition, per the corrections policy § 5.
6. Log entry in the public removals log.

5. The removals log

We publish a running log of removals. Each entry records the edition, the date, the ground category, and the scope of what was removed. It does not record the identity of the subject or the requester, and it does not restate the removed material — a log that re-identifies the person it protected would defeat itself.

The log exists so that the removal power is visible and countable. A publisher who can quietly remove material is a publisher whose archive cannot be trusted; the log is what makes our removals auditable without exposing the people they protect.

5a. What a reader following a citation finds

The log above is centralised and answers “what has been removed”. It does not answer the question a specific reader has: they followed a citation to a stable public ID and got nothing.

A stable ID that returns nothing is indistinguishable from a publication failure, a bad link, or a failed site move. A researcher who cited an item in their own work cannot tell whether we withdrew it or whether publication failed — and the second explanation is the one they will assume, because it is the more common cause.

Every removal therefore leaves a marker at the item’s own stable ID, recording that the item was withdrawn and the date. Nothing else.

• It does not record the ground category. The centralised log carries that, aggregated, where it cannot be tied back to a specific subject by anyone reading a single URL.
• It does not record the subject or requester, for the same reason § 5 gives.

The distinction matters: the log makes the removal power auditable, and the marker makes an individual removal legible to the person who cited it. Those are different jobs and one artifact cannot do both — a per-item marker detailed enough to be auditable would re-identify the subject, and a log aggregated enough to protect the subject cannot be found by someone holding a dead link.

6. The limit we cannot engineer around — read this before relying on removal

Our editions are published under CC BY 4.0 and are explicitly designed to be copied, mirrored, and ingested by automated research agents. The Parallax Research Atlas exists to be consumed programmatically.

Removal from our surface does not retract copies. We can remove material from the current edition and every future one. We cannot remove it from a mirror, a web archive, a downstream dataset, a research agent’s index, or anyone’s local copy. No publisher operating under an open licence can, and we will not imply otherwise to make a removal feel more complete than it is.

What this means practically:

• Prevention carries the weight, not removal. The pre-publication gates (methods.md) are the real protection for victims and minors. The removal route is a backstop for what those gates miss.
• We will tell a requester honestly what removal achieves and what it does not, at the point they ask, rather than after.
• For grounds 1 and 2 we will, on request and where we can identify them, notify known downstream mirrors. We cannot compel them.

7. Escalation

If a requester disagrees with a decision, they may ask for it to be reconsidered in writing. They should know that this goes back to the same person. The publication owner and the release approver are the same individual for the 2026-10 edition, so there is no independent internal appeal. We would rather state that than describe an escalation route that does not exist.