Skip to content

Privacy and preparation

Dataset incident versus quality defect: when an error needs escalation

By SourceX Editorial · Reviewed by Noah Loul ·

10 sections

Short answer

Escalate a dataset error as an incident when it exposes personal or confidential information, puts records outside the agreed scope into someone's hands, or shows an unexplained change to the data. Wrong labels, duplicates and missing values are quality defects for the repair queue, unless they have reached recipients in a way a routine fix cannot undo. When unsure, open the incident, preserve the evidence and downgrade later.

Record timeline

Dataset incident

This page covers dataset incident versus quality defect: when an error needs escalation.Illustrative

Key takeaways

  • A quality defect fails the dataset's own specification; an incident has consequences outside it, such as exposure, scope breach or unexplained change.
  • Route by consequence, not row count: one exposed record outweighs thousands of wrong labels.
  • Privacy notification clocks can run from awareness, so record the discovery time and escalate the same day.
  • Preserve the affected version, manifest, recipient log and failed-check output before any repair or rerun.
  • Exposure and scope checks should fail a release on the first hit; label and duplicate checks can use thresholds.

When is a dataset error an incident rather than a quality defect?

A quality defect means the data fails its own specification: a label is wrong, a field is empty, a record appears twice. An incident means the error has consequences outside the dataset: someone may have received information they should not have, an agreement or law may have been breached, or the data changed in a way nobody can explain. The first is fixed in the next release. The second needs owners, a clock and evidence before anyone touches the file.

Security practice draws the line in the same place. NIST's incident response guidance, SP 800-61 Revision 3, defines a cybersecurity incident as an occurrence that actually or imminently jeopardizes, without lawful authority, the integrity, confidentiality or availability of information or an information system, or that violates or imminently threatens to violate law or security policies. Most dataset errors touch integrity in a loose sense; the triage below asks whether they touch it in that stronger one.

This page gives a triage decision guide for teams that prepare and license business records. It is general information to be adapted with counsel and the security owner, who should review the guide before it is adopted, not a substitute for their judgment on a live case.

The dataset incident versus quality defect triage guide

Ask four questions in order and stop at the first yes. Each yes routes the error to an incident with a named decision owner; four noes make it a quality defect. The questions are deliberately about consequences, not about how many rows are wrong, because a single exposed record can matter more than ten thousand bad labels.

The dataset incident versus quality defect triage guide
QuestionExamples of a yesRoute if yesDecision owner
Does the error reveal personal, confidential or secret information?A free-text field skipped redaction; an API token sits in a log line; a customer's name survives in a file namePrivacy or security incidentPrivacy lead with counsel and security
Does the release contain records outside the agreed scope?Records from an excluded business unit; records withdrawn before the release; a field the agreement excludedRights incidentContract owner with counsel
Is there an unexplained change to source or released data?Rows altered with no matching pipeline run; checksums differ from the release manifestIntegrity incidentSecurity owner
Has a defect already reached recipients in a way that changes their results?A mislabeled evaluation split a recipient has already reported results onEscalated defect with noticeDataset operations owner
None of the aboveDuplicates, wrong labels, missing values found before or after releaseQuality defectDataset operations owner

Which severity class applies, and what does each one start?

Severity sets who is told and how fast, so define it before the first error rather than during one. The classes below separate incidents from defects and give each a first action; adapt the timing to your agreements and to counsel's advice.

Which severity class applies, and what does each one start?
ClassTypical triggerFirst actionCloses when
I-1 exposurePersonal or confidential data in a released or shared fileStop distribution, record discovery time, notify counsel and security the same dayCounsel signs off on notifications and recipients confirm deletion of the affected version
I-2 scopeOut-of-scope or withdrawn records in a releaseStop distribution, notify the contract owner, list affected recipientsCorrected release delivered and old version deletion confirmed
I-3 integrityUnexplained change to source or released dataFreeze the pipeline and preserve logs before any rerunCause found and a clean rebuild verified against the manifest
D-1 material defectDefect that changes results a recipient relies onOpen a tracked issue and send recipients a change noticeRepaired release delivered with regression checks passed
D-2 routine defectDefect below agreed thresholds or caught before releaseLog it in the repair queueFix ships in the next release with a release note

Why does the first hour of an exposure matter so much?

Because notification clocks in privacy law often run from awareness, not from the time the cause is understood. Under the EU General Data Protection Regulation, a personal data breach is a breach of security leading to the accidental or unlawful destruction, loss, alteration, unauthorised disclosure of, or access to, personal data. Accidental disclosure counts. Article 33 then requires a controller to notify the supervisory authority without undue delay and, where feasible, within 72 hours of becoming aware, unless the breach is unlikely to result in a risk to people's rights and freedoms, and requires a processor to inform the controller without undue delay.

United States health privacy rules take a similar stance for protected health information. HHS explains that an impermissible use or disclosure is presumed to be a breach unless a risk assessment shows a low probability that the information has been compromised, weighing four factors: the nature and extent of the information, who received it, whether it was actually acquired or viewed, and how far the risk has been mitigated.

Whether either rule applies, and who must notify whom, depends on the parties' roles, the data and the agreements, and that is counsel's call. The data team's job is to make that call possible: record the moment of discovery, keep the affected file and its recipient list intact, and escalate the same day.

What evidence has to be preserved before anyone repairs the file?

Everything a reviewer will need to reconstruct what was released, to whom and why. Repairs overwrite the very evidence an incident review depends on, and some tools do it automatically. dbt, for example, documents that when a data test stores its failing records, each new run replaces the previous failures for that test, even when the new run finds none. If a failed check is your first evidence, copy its output somewhere it will not be overwritten before rerunning anything.

  • Issue ID and severity class, with the time of discovery and who discovered it.
  • The affected release version, its manifest and checksums, preserved unchanged.
  • Links from each affected released record back to its source records and the pipeline run that produced it.
  • The recipient and version log: who received the affected version, when and by what channel.
  • The repair: the code or configuration change, who reviewed it and the regression checks it passed.
  • The change notice sent to recipients and their acknowledgments.
  • Reconciliation: confirmation that superseded copies were deleted or quarantined where the agreement requires it.

Which released examples and recipients does a source change affect?

Answer this from lineage, not from memory. Start at the faulty source records, follow them to every derived file and release that included them, then to every recipient of those releases. A release manifest that lists source record keys per output file turns this into a query; without one it becomes a search through folders and email.

The Datasheets for Datasets framework asks dataset creators to say whether there is an erratum and whether the dataset will be updated to correct labeling errors, add instances or delete instances, including how often, by whom and how updates will be communicated. Answering those questions before the first release is what makes the affected-recipient list quick to produce later.

Keep the answer per version. A recipient who already moved to a later release may still hold the affected one, and a recipient who used the affected version to train or evaluate a model may need more than a replacement file; what the agreement requires in that case is for the contract owner and counsel to decide.

Where should an automated check warn, fail the release or open an incident?

Map each check to the triage guide instead of giving every check the same severity. dbt's test configuration shows the mechanics: a test can be set to warn or to error, and error_if and warn_if thresholds make the result depend on the number of failures, so one duplicate can warn while ten fail the run.

Use counts for defects and zero tolerance for exposures. A check for duplicate tickets can warn below an agreed threshold and fail the release above it. A check that finds a phone number or an email address in a field that should be redacted should fail the release on the first hit and open an I-1 review automatically, because the count of exposed records matters less than the fact that the redaction step failed.

  • Redaction and secret-pattern checks: fail on any hit; open an incident review.
  • Scope checks against the agreed inclusion list: fail on any hit; route to the contract owner.
  • Manifest and checksum checks on released files: fail on any mismatch; route to security.
  • Label, duplicate and completeness checks: warn or fail by threshold; route to the repair queue.

Worked example (illustrative): a note field that bypassed redaction

This case is invented to show the triage; it describes no real company or dataset. A company prepares a support-ticket dataset of 48,000 tickets and has delivered version 3 to two recipients. A schema change in the help desk export added an internal notes column, and the pipeline passed it through without running the redaction step that cleans the public reply text.

During a routine scan, an analyst finds phone numbers in 120 rows of the new column. Question one is a yes, so this is an I-1 exposure. Distribution stops, the discovery time is logged, and counsel and the security owner are told the same day. The version 3 files, manifest and recipient log are preserved untouched, and the scan output is copied before any rerun.

The same scan also finds 300 duplicate tickets created by a retry in the export job. On their own those would be a D-2 routine defect. They are fixed in the same corrected release, but they are logged as a separate issue so the incident record stays limited to the exposure.

The repair routes every free-text column through redaction by default and adds a check that fails the release when an unredacted column appears. Version 3.1 ships after regression checks pass, recipients receive a change notice, and the incident closes only when counsel has decided on notifications and both recipients have confirmed deletion of version 3.

When should a routine defect still be escalated, and where does this guide stop?

Escalate a defect when its pattern, not its size, is the problem. The same defect class returning in three releases suggests a control has failed, and that deserves an owner above the repair queue. A defect in a field the agreement specifically describes, such as a promised completeness level, can become a contract issue even when no data was exposed.

The guide has limits. It cannot decide whether a particular disclosure is reportable, whether a recipient must delete a trained model, or which party holds which legal role; those depend on facts, agreements and jurisdiction. It also assumes you can detect the error. Exposure checks only find the patterns they look for, so review new fields by hand when a source schema changes.

How incident triage fits a data licensing agreement

Write the triage classes, owners and notice steps into the delivery plan before the first release, so both sides already agree what a change notice and a deletion confirmation look like. That is easier when scope is written down, because the scope question in the guide needs an agreed inclusion list to test against.

When SourceX is involved, no records reach it until the supplying company has approved scope and terms and signed, and de-identification runs through approved processing routes managed by SourceX. That makes the signed agreement the natural home for the triage owners on each side, rather than a later email.

Frequently asked questions

Is a wrong label ever an incident?

Usually not. A wrong label is a quality defect handled through repair and a release note. It becomes an escalated defect when it changes results a recipient has already relied on, and part of an incident only if it comes with exposure, out-of-scope records or an unexplained change.

Does an accidental disclosure count as a breach under the GDPR?

The GDPR definition of a personal data breach includes accidental as well as unlawful disclosure of personal data. Whether a specific event must be notified depends on the risk to individuals and the parties' roles, so counsel should assess it promptly. Record the time your team became aware.

Who should own the decision to call something an incident?

Name the owner per route before anything goes wrong: a privacy lead with counsel for exposures, the contract owner for scope errors, the security owner for unexplained changes. The dataset operations owner runs routine defects. Anyone should be able to open an incident; only the named owner downgrades it.

Should we fix the data first and investigate afterward?

No. Stop distribution first, then preserve the affected version, its manifest, the recipient log and any failed check output. Some tools overwrite earlier test failures on the next run, so a quick rerun can erase your first evidence.

What should a change notice to recipients include?

The affected version, what was wrong in plain terms, which records or files are affected, the corrected version and its identifier, and what the recipient is asked to do with the old copy. Keep the tone factual and send it through the channel the agreement names.

Sources

  • NIST SP 800-61 Revision 3 defines a cybersecurity incident as an occurrence that actually or imminently jeopardizes, without lawful authority, the integrity, confidentiality or availability of information or an information system, or violates or imminently threatens to violate law or security policies. Source
  • The GDPR defines a personal data breach to include accidental or unlawful destruction, loss, alteration, unauthorised disclosure of, or access to, personal data, and Article 33 sets the controller's notification duty of without undue delay and, where feasible, within 72 hours unless the breach is unlikely to result in a risk. Source
  • HHS explains that an impermissible use or disclosure of protected health information is presumed to be a breach unless a four-factor risk assessment shows a low probability of compromise. Source
  • dbt data tests can be configured with severity warn or error and with error_if and warn_if thresholds on the number of failures. Source
  • When dbt stores test failures, a test's results always replace previous failures for the same test, even when the new run has no failures. Source
  • Datasheets for Datasets asks whether there is an erratum and whether the dataset will be updated to correct labeling errors, add or delete instances, and how updates will be communicated. Source

See if your company qualifies

A short company assessment. No data uploads are needed.

See if your company qualifies