Skip to content

Getting started

Supplier dataset scope revision: six upgrades and what each reopens

By SourceX Editorial · Reviewed by Noah Loul ·

10 sections

Short answer

A supplier dataset scope revision is due when a company holds more or better records than its application described: cleaner data, longer history, more units, new fields, a second system or a new record type. Do not let the extra data ride along at delivery. Classify the upgrade, issue a new manifest version, rerun only the reviews it touches and have the same approvers sign again. More data does not promise a better offer.

Process

Supplier dataset scope

This page covers supplier dataset scope revision: six upgrades and what each reopens.Illustrative

Key takeaways

  • Every approval was given against a description, so data the description does not name has not been approved, however good it is.
  • Six upgrade types reopen different reviews: cleaner records mostly reopen quality, while a new record type such as voice notes reopens rights, privacy, quality and buyer fit.
  • Croissant's convention calls added records, fields or record sets a minor version and edited existing data a major one, but the version step does not measure review burden.
  • The EU AI Act's mandatory training-content summary, applicable since 2 August 2025, means buyers building general-purpose models need an accurate list of what they received.
  • Offer the upgrade as a separately scoped option; the buyer may want only the slice it first evaluated.

When does better data call for a supplier dataset scope revision?

Revise the scope whenever the dataset you could deliver differs from the one your application described in its systems, record types, fields, date range, business units or legal entities. Every review so far, from qualification to rights and privacy screening, was done against that description, so anything outside it has not been reviewed, however valuable it looks.

Companies find better data for ordinary reasons. The person who filled in the application knew the help desk but not the billing system; an archive from an old migration turns up more years; an engineer finds that resolution codes, which would let a buyer grade outcomes, sat in a field nobody exported. The temptation is to fold the extra material into the delivery as a pleasant surprise.

A supplier dataset scope revision is the controlled alternative: a new version of the description, a record of what changed, a rerun of only the reviews the change touches and a fresh sign-off from the same approvers. The evidence behind a good decision is concrete: the original application text, an inventory of the new material by system and date range, the contracts and notices that govern it, and a sample showing its quality.

Six kinds of upgrade, and why they are not equal

Sort the discovery into one of six types before estimating anything, because each type reopens a different set of reviews. Two upgrades of the same size can carry very different risk: four more years of the same tickets is mostly a volume and consistency question, while a few thousand voice notes raise a privacy review the original scope never needed.

The six types, each shown for an illustrative field service company that applied with three years of service tickets from one system, are listed below:

  • Cleaner records: the same tickets, deduplicated and with corrected timestamps.
  • Longer history: seven years of the same tickets instead of three, recovered from a migration archive.
  • More units: tickets from a second region or brand, which may sit in a different legal entity.
  • New fields: resolution codes and parts used, stored in the same system but never exported.
  • A second system: invoices from the accounting system, linked to tickets by job number.
  • A new record type: technician voice notes, photos or free-text customer messages.

The scope revision worksheet: what each upgrade reopens

For each upgrade type, the worksheet says what to change in the application and manifest, which reviews to rerun, the version step a standard convention would assign and how much internal effort to expect. Fill it in before telling anyone the dataset got bigger, so the news arrives with its consequences attached.

The version column follows the Croissant dataset format: a minor version adds records, fields or record sets while existing data stays retrievable as before, and a major version edits, removes or reshuffles existing data. Effort bands are editable planning assumptions, not measurements. The figure shows the field service example as a before and after manifest.

The scope revision worksheet: what each upgrade reopens
UpgradeUpdate in the application and manifestReviews to rerunVersion step (Croissant convention)Effort band (editable)
Cleaner recordsChange log of fixes, record counts before and after, and the method usedQuality; privacy only if a fix touched identifiersMajor: existing data edited or removedLow
Longer historyNew date range, volume band and any period with different source rulesRights and privacy for the added years; quality sample from the oldest periodMinor: records addedMedium
More unitsBusiness units, the legal entity holding each set, and volume by unitRights for each entity; privacy where notices differ by unitMinor: records addedMedium
New fieldsField list with definitions, null rates and source tablesPrivacy and re-identification; quality of the new fieldsMinor: fields addedMedium
A second systemSystem, owner, join key, linkage rate and date rangeRights under that system's vendor and customer contracts; privacy; quality of the joinMinor: record set addedHigh
A new record typeRecord type, format, volume and what it reveals about peopleRights, privacy, quality and buyer fit, each in fullMinor: record set addedHigh
A before and after manifest: three years of service tickets from one system as described, and a revised version adding seven years of tickets, resolution codes and parts, invoices and technician voice notes, each marked with the reviews it reopens.
Illustrative field service example: what the revision adds and which reviews each addition reopens.

The version number does not measure the review burden

A version step tells a data user whether earlier work still runs; it says nothing about how much the owner must recheck. In the worksheet, the heaviest review sits on a minor version, a new record type, while the only major version, cleaner records, usually needs the lightest review.

Croissant, the machine-readable dataset format published by MLCommons, recommends MAJOR.MINOR.PATCH numbers: a patch for the same data packaged differently, a minor version when existing data is unchanged and still retrievable but more has been added, and a major version when existing data has been edited, removed or shuffled. Version 1.1 of the specification, published in January 2026, also keeps a separate version number for the metadata, so a corrected description can change without implying the files changed, and it strongly recommends SHA-256 checksums for the files of a versioned dataset.

In the field service example, every addition leaves the original three years of tickets retrievable, so the description moves from version 1.0.0 to 1.1.0, even though the voice notes alone reopen all four reviews. Use both signals. The version number tells the buyer what to retest; the reviews column tells your approvers what to sign again. Approvers who see only a minor bump may assume nothing important changed, which is exactly the wrong conclusion for a new record type.

What the revised manifest has to say about the new material

The revised manifest keeps the original description as the first version and adds the new material as a dated, attributed change, so anyone can see what was first offered and what was added. Describe the new material with the same fields as the old, not in a sentence appended to the end.

Datasheets for Datasets, the documentation framework proposed by Timnit Gebru and colleagues, groups its questions by stage of a dataset's life, from motivation and composition through collection, preprocessing, uses and distribution to maintenance. One composition question asks whether a dataset contains all possible instances or a sample from a larger set. Answer it for the new material as well: are the seven years complete, or only what survived a migration?

The Data and Trust Alliance's Data Provenance Standards organize 22 metadata fields into three standards, source, provenance and use, drawn from use cases across 15 industries, and the alliance reports that IBM tested them in early 2024 in its clearance process for datasets used to train foundation models. Those three headings make a practical check for each manifest line: where the records came from, how and when they were generated, and which uses and restrictions apply.

Which reviews the new material reopens, and why buyers care

New material reopens rights and privacy whenever it comes from a different system, entity or period, because the contracts and notices that govern it may differ from the ones already checked. Quality and buyer fit reopen whenever the new material changes what a buyer would train or test on.

Rights first. Records in a second system can be bound by that vendor's terms and by the customer contracts of a different product line, and content other people created, such as documents customers uploaded, may not be yours to license at all; OpenAI's data partnerships page, for one, rules out information that belongs to someone else. Privacy next. Older records were collected under the notices of their day, and in February 2024 FTC staff cautioned that moving to more permissive data practices, such as using consumers' data for AI training, while disclosing that only through a surreptitious, retroactive change to terms may be unfair or deceptive.

Buyers need the revision for their own reasons. Under the EU AI Act, providers of general-purpose AI models must publish a summary of their training content using the European Commission's mandatory template, which lists data sources by type, private datasets among them. The obligation has applied since 2 August 2025, with enforcement by the AI Office from 2 August 2026. A buyer that received records its own documentation does not list has a disclosure gap it never chose.

Why silent expansion fails even when the extra data is good

Silent expansion fails because the agreement, the approvals and the dataset documentation end up describing a smaller dataset than the one delivered. Each gap is small on the day of delivery and expensive later.

The approvers signed for the described scope, so the extra records leave without anyone having approved them. Acceptance testing runs against the agreed description, so the buyer may reject or quarantine records it did not expect. And documentation drifts. Reviewing more than 1,800 public text datasets, the Data Provenance Initiative reported licenses missing from over 70 percent of listings on popular hosting sites and error rates beyond 50 percent, and it found license information often dropped away when datasets were merged into larger collections. Records that arrive without a matching description are the ones most likely to lose their terms downstream.

The rule that follows is short: deliver nothing the signed scope does not name. If the upgrade is worth having, it is worth a new version and a new approval.

Does a better dataset mean a better offer?

Not necessarily, and assuming it does is the main reason companies skip the revision. A buyer evaluates records against its own task, so more records help only if they serve that task, and extra material can add review time and risk that the buyer weighs against it.

On volume, SourceX publishes only a general guide: buyers tend to want a single system holding tens of thousands of records or more, kept consistently over a span of years, usually at least two, and a smaller set can still qualify when the work it records is rare. Every case is judged deal by deal. A longer history can move a borderline dataset closer to what buyers usually ask for, but it cannot change a buyer's task or budget. Eligibility, an assessment or an indicative figure does not mean the records will be selected, that a deal will close or that a particular amount will be paid.

Present the upgrade as an option rather than a replacement. The originally described scope stays on the table, and the additions are scoped, versioned and reviewed separately, so the buyer can take them, take some or leave them.

The scope revision memo: decision, limits and next evidence

Close the revision with a one-page memo that the original approvers sign, so the decision sits next to its evidence. The memo replaces nothing; it rests on top of the original application and the earlier approvals.

At a glance

The scope revision memo: decision, limits and next evidence

State its limits plainly: it covers the material inventoried on the date written and the reviews listed, and any later discovery starts a new revision. Use the headings listed below:

  • Decision: offer the upgrade as an option, fold it into the scope, defer it or exclude it.
  • Upgrade type from the worksheet, with the systems, date ranges and legal entities involved.
  • Manifest version before and after, with the change log.
  • Reviews rerun: who ran each one, the result and any restriction it added.
  • Effort: the internal work the upgrade adds before anything can be delivered.
  • Effect on the buyer conversation: what changes in the description the buyer receives.
  • Next evidence needed, such as a quality sample from the oldest years.
  • Approvers, with the date each one signed.

Revising your scope before a SourceX agreement

Revise the description, not the delivery. The first step with SourceX is a short company assessment that needs no data upload, so an upgrade found early changes what you describe rather than files you have sent. Delivery waits on two acts by the company, approving the scope and terms and signing the agreement, and that approval is the moment to make sure the signed scope names everything you mean to license.

SourceX's role spans the licensing path for operational records sold to AI developers, from sourcing, qualifying and rights-checking the data to contracting, preparation, delivery and payment. If better data surfaces after signing, the signed agreement still defines what is licensed, so the additions wait for written agreement and the same reviews. An assessment is not an offer, and an updated scope does not mean selection, a closing or any amount being paid.

Frequently asked questions

Should a supplier tell the buyer it has more data than it described?

Yes, as a documented option rather than a surprise in the delivery. Send a revised manifest that keeps the original scope and lists the additions separately, with the reviews each one passed, so the buyer can take the additions, part of them or none.

Do extra years of the same records need a new rights review?

Usually a partial one. The added years may have been collected under older customer contracts, notices or vendor terms, so check what governed each period; if the same terms applied throughout, the review is short. Take a quality sample from the oldest period too, where gaps are common.

What version number should a revised dataset description get?

Under the Croissant convention, added records, fields or record sets that leave existing data retrievable are a minor version, while edits or removals of existing data are a major version. Croissant also keeps a separate metadata version, which suits a correction to the description alone.

Will a bigger dataset raise the offer?

Not necessarily. Value depends on the buyer's task, and more material can add review time and risk. An assessment or indicative figure is not an offer, and an updated scope does not mean selection, a closing or payment.

What if the better data turns up after signing?

Do not ship it as a larger delivery. The signed agreement defines what is licensed, so the additions need the parties' written agreement and the same rights, privacy and quality reviews they would have had before signing.

Sources

  • Croissant 1.1 (published January 29, 2026) recommends MAJOR.MINOR.PATCH dataset versions: minor when existing data is unchanged but data is added, major when existing data is edited, removed or shuffled; it keeps a separate metadata version and strongly recommends SHA-256 file checksums. Source
  • Datasheets for Datasets organizes documentation questions by lifecycle stage, from motivation and composition to maintenance, including whether a dataset contains all possible instances or a sample. Source
  • The Data and Trust Alliance's Data Provenance Standards organize 22 metadata fields into source, provenance and use standards drawn from 15 industries; IBM tested them in early 2024 in dataset clearance. Source
  • The Commission's template for general-purpose AI model providers' public training content summary is mandatory under Article 53(1)(d), lists data sources by type including private datasets, applies from 2 August 2025 and is enforceable by the AI Office from 2 August 2026. Source
  • FTC technology staff wrote in February 2024 that adopting more permissive data practices, such as AI training, and disclosing them only through a surreptitious, retroactive change to terms may be unfair or deceptive. Source
  • OpenAI's data partnerships page says it is not seeking datasets with information that belongs to a third party. Source
  • The Data Provenance Initiative audited more than 1,800 text datasets, found license omission above 70 percent and errors above 50 percent on hosting sites, and found license information often lost when datasets were aggregated. Source

See if your company qualifies

A short company assessment. No data uploads are needed.

See if your company qualifies