Skip to content

Privacy and preparation

Dataset supplier change request: when one more field reopens scope

By SourceX Editorial · Reviewed by Noah Loul ·

10 sections

Short answer

Reopen scope review whenever a requested field is not already named in the signed scope, comes from a different system or owner, needs a new transformation, or could make records easier to tie to a person when combined with fields already delivered. Only re-packaging approved fields or correcting an error in them skips the full review, and even those get a new version and a notice to every recipient.

License terms

Dataset supplier change

This page covers dataset supplier change request: when one more field reopens scope.Illustrative

Key takeaways

  • A requested field outside the signed scope is a new decision, not an export tweak.
  • Four tests decide how far review reopens: scope, source system, transformation and re-identification risk.
  • Fields that look harmless alone can narrow records toward individuals when combined with what was already delivered.
  • New fields ship as a new minor version; corrections and removals change existing data and ship as a major version.
  • Keep a recipient and version log so corrections reach everyone who received the affected data.
  • Narrowing a field often meets the buyer's purpose with far less exposure than declining.

When should a buyer's request for one more field reopen scope review?

Reopen it whenever the field is not already in the approved scope. A dataset supplier change request is a new decision about rights, privacy, quality and delivery, not a column added to an export job. Four conditions decide how far the review goes: whether the field is named in the signed scope schedule, whether it comes from the same system and owner as the approved fields, whether it needs a new transformation, and whether it changes what the combined record can reveal about a person.

If none of those apply, for example the buyer wants the same approved fields as Parquet instead of CSV, a light review is enough. If any one applies, the change goes back through the same gates the original dataset passed. The rest of this page gives a triage table, a change request record, an approval workflow and a worked example a data program can use as is.

Why is one more field never just one more column?

Because each of the original approvals was made for a specific set of fields. The rights review checked who owns the source tables and whether the company's customer and vendor terms allow licensing them. The privacy review weighed what the selected fields reveal. The quality review measured coverage across the date range. A new field can break any of those conclusions without anyone noticing, because the export still runs.

Privacy is the easiest to miss. EU data protection law requires personal data to be adequate, relevant and limited to what is necessary for the purpose, so a field added because it might be useful is a weaker case than one tied to a stated purpose. Re-identification risk also grows with combinations: NIST's survey of de-identification, NISTIR 8053, describes quasi-identifiers as attributes that do not identify a person alone but can be combined and linked with other data. A technician's home region looks harmless; next to job dates and a small customer list it may not be.

Delivery is the last gate. The Croissant metadata format for machine learning datasets recommends semantic versioning and says new fields belong in a minor version, while edits to existing data call for a major version. A change that skips review usually also skips the version bump, and then nobody can say which training run used which shape of the data.

Triage table: which change requests reopen which reviews

Classify every request before anyone estimates effort. The version column follows the Croissant recommendation; the approver column is a reasonable default for a company licensing its records and should be adjusted to your own sign-off rules.

Triage table: which change requests reopen which reviews
Change typeExampleReviews to reopenVersion effectWho approves
Re-package approved fieldsSame fields delivered as Parquet instead of CSVDelivery check onlyPatchDataset owner
Correct an error in a delivered fieldTimestamps shipped in the wrong time zoneQuality, then recipient noticeMajor, because existing data changedDataset owner and quality reviewer
Add a field derived only from approved fieldsTicket handle time computed from approved open and close timesQuality and documentationMinorDataset owner
Add a field from the same system, not in scopePriority field on the same support ticketsRights, privacy screen, qualityMinor, after amendmentData owner, privacy reviewer, signatory
Add a field from a different systemCustomer industry from the CRM joined to ticketsFull review: rights, privacy, quality, join methodMinor, after amendmentBoth system owners, privacy reviewer, signatory
Add free textInternal agent notes on each ticketFull review plus text de-identification testMinor, after amendmentPrivacy reviewer and signatory
Add a likely quasi-identifierExact service date, ZIP code or job titlePrivacy and re-identification review firstMinor, after amendment, or declinedPrivacy reviewer, counsel if needed, signatory
Extend the date range or add recordsTwo more years of the same ticketsRights for the new period, qualityMinor if old records are unchangedData owner and signatory
Remove a fieldDrop a field found to hold personal dataPrivacy, then recipient noticeMajorPrivacy reviewer and dataset owner

What goes in the change request record before anyone says yes?

One record per request, filled in order, with nothing built until the decision field is complete. Keeping it as a row in the same tracker as the original scope review makes it easy to show later why a field was added and who approved it.

  • Request ID, date received and the buyer's reference.
  • Requested field, in the buyer's words and in the source system's field name.
  • Stated purpose: what the buyer will use the field for, in one sentence.
  • Source system, table and data owner; flag when the owner differs from the original scope.
  • Triage class from the table above.
  • Scope check: the clause or schedule line that covers the field, or 'not covered'.
  • Privacy screen: whether the field is personal data, free text or a likely quasi-identifier, and the result of a combination check with already delivered fields.
  • Sample result: coverage and null rate across the full date range, and any period where the field was used differently.
  • Transformation: the exact rule for derivation, redaction or generalization.
  • Affected releases and recipients: every version and recipient the change touches.
  • Version proposal and draft release note.
  • Decision (approved, approved with changes, declined), approvers and date.
  • Change notice text, sent date and acknowledgment.

The approval workflow, from intake to release

The workflow mirrors standard configuration change control. NIST's guide to security-focused configuration management, SP 800-128, describes changes as formally identified, proposed, reviewed, analyzed for security impact, tested and approved before implementation. The same order works for dataset changes.

Steps

The approval workflow, from intake to release

  • Intake: log the request and acknowledge it; promise a decision date, not a delivery date.
  • Triage: assign the change type and the reviews it reopens.
  • Scope check: compare the field against the signed scope schedule. If it is not covered, the request needs an amendment before any build.
  • Privacy and re-identification review: test the field alone and in combination with delivered fields; generalize or decline when the combination narrows records toward individuals.
  • Sample: pull the field for a small slice across the full date range and record coverage, null rates and any change in meaning over time.
  • Decision and signature: approve, approve with a transformation, or decline, and amend the agreement where needed.
  • Build and regression check: produce the new version from scratch, confirm previously delivered fields are byte-identical where they should be, and recompute checksums.
  • Release: publish a new version with release notes. Semantic Versioning's rule that a released version's contents must not be modified applies here; corrections ship as a new version.
  • Record: update the recipient and version log and close the request.

Worked example: a buyer asks for technician region on closed service tickets

This example is illustrative. A field service company has licensed closed service tickets with job type, symptom, parts used, resolution code and open and close timestamps, and delivered it as version 1.0.0. The buyer now asks for the technician's region to study regional differences in first-time fix.

Triage puts it in the same-system, not-in-scope class: the region sits on the technician record in the same dispatch system, but the scope schedule lists only ticket fields. Scope check: not covered, so an amendment is needed. Privacy screen: region alone is coarse, but the company has some regions with very few technicians, and combined with timestamps a region could point to one person's work history. The reviewer approves the field generalized to a larger area, with regions under an illustrative minimum of five technicians merged into a neighbor.

The sample shows the region field was reorganized partway through the date range, so the transformation maps old codes to new ones and the release note says so. After the amendment is signed, the team builds version 1.1.0, confirms the original fields are unchanged, and logs the release for the one recipient of 1.0.0.

Change one detail and the answer changes. Had the buyer asked for the technician ID instead of the region, the request would move to the likely quasi-identifier row: a stable ID links every job one employee handled, which is a work history, and the company's employee notices may not cover sharing it. The reviewer would most likely decline the ID and offer the generalized region or a per-release random key that cannot be joined across versions, and would record that choice so the next request starts from it.

Which released versions and recipients does the change touch?

Every recipient of every version that contained the affected data, plus anyone who received a derived subset. Keep a log that ties each release to its recipients, delivery dates and the agreement that covers it, because adding a field touches only future versions while correcting or removing one reaches back.

For corrections, the obligation can be legal as well as contractual. Article 19 of the GDPR requires a controller to communicate any rectification or erasure of personal data to each recipient to whom the data were disclosed, unless that proves impossible or involves disproportionate effort. The Datasheets for Datasets questions make the same point for any dataset: say whether it will be updated to correct errors or add and remove instances, how often, by whom, and how updates will be communicated.

  • Release ID, version number and checksum.
  • Recipient, delivery date and delivery method.
  • Agreement and scope schedule version in force at delivery.
  • Derived subsets or samples sent from that release.
  • Notices sent about the release, with dates and acknowledgments.

Which changes can take a fast path, and which cannot?

Two kinds of change should move faster than the full workflow, but neither skips the record. The first is a correction to a field already in scope, such as a time zone error. The data and the rights do not change, so the scope check is quick, but existing values do change, which makes it a major version and triggers notices to every recipient.

The second is a removal for privacy reasons. If a delivered field turns out to contain personal data that should not have been there, withdraw first and review after: stop further deliveries, notify recipients and ask them to replace the affected version, then run the review to find how the field slipped through.

The opposite is never fast: a new field from a different system, free text, or anything near an identifier. Those wait for the full review, however small the request sounds.

Signals that a field request needs counsel or a new agreement

Stop and escalate when the requested field belongs to a system the original rights review did not cover, when it describes employees or customers in a way the company's notices may not have anticipated, or when the buyer's stated purpose differs from the purpose in the agreement. Escalate as well when the field is regulated in your sector, for example health or financial account information, or when the requester cannot say what the field is for.

Narrowing is often the better answer than declining. A coarser region, a banded date, a coded category instead of free text, or a field limited to recent years can meet the buyer's purpose with far less exposure. Record the narrowed version as the approved transformation so the next request starts from it.

Field requests on a licensed dataset: who signs off

When a dataset is licensed through SourceX, the company approves the scope and terms and signs before data is delivered, so a field outside the signed scope is not part of the licensed dataset unless the company approves a change. For businesses licensing operational records directly to SourceX, SourceX covers rights review, contracting and preparation. A company that has not started yet can check fit with a short assessment that needs no data upload.

Frequently asked questions

Does adding a field always require a contract amendment?

Not always, but it does whenever the field is outside the signed scope schedule. Adding a field derived only from approved fields, or re-packaging approved fields in a new format, usually stays within scope. Counsel should confirm the reading of your specific agreement.

What version number should a dataset get after a new field is added?

The Croissant format recommends semantic versioning and treats new fields as a minor version, as long as the existing data can still be retrieved unchanged. Edits to existing records, including corrections, call for a major version. Keep the previous version available for reproducibility where the agreement allows.

Who should approve a buyer's change request?

The owner of the source system, a privacy reviewer for anything that touches people, and whoever signs agreements for the company when scope changes. Format-only changes can be approved by the dataset owner alone. Write the approvers into the change request record.

How do you check whether a new field raises re-identification risk?

Test it alone and in combination with fields already delivered. Fields such as exact dates, small-area locations and job titles act as quasi-identifiers when combined, so count how many records share each combination and generalize or decline where groups are small.

Do recipients of older versions need to hear about a new field?

Not for an addition, which affects only new versions, though a release note helps. They do need a notice for corrections and removals, because their copy now differs from the corrected data. Under the GDPR, rectification or erasure of personal data must be communicated to each recipient unless that is impossible or disproportionate.

Sources

  • The Croissant specification recommends semantic versioning for datasets: patch for re-serialization, minor for added data such as new fields, major when existing data changes. Source
  • Semantic Versioning says the contents of a released version must not be modified; changes ship as a new version. Source
  • GDPR Article 5(1)(c) requires personal data to be adequate, relevant and limited to what is necessary for the purposes. Source
  • GDPR Article 19 requires rectification or erasure to be communicated to each recipient unless impossible or disproportionate. Source
  • NISTIR 8053 describes quasi-identifiers as attributes that do not identify a person alone but can be combined and linked with other data. Source
  • NIST SP 800-128 describes changes being formally identified, proposed, reviewed, analyzed for security impact, tested and approved before implementation. Source
  • Datasheets for Datasets asks whether a dataset will be updated, how often, by whom 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