Skip to content

Rights and contracts

Portfolio data room dataset evidence: an index and version convention

By SourceX Editorial · Reviewed by Noah Loul ·

9 sections

Short answer

Build a dataset evidence folder for buyer diligence around one index: each document gets a stable ID, the entity and system it covers, the dataset versions it supports, a SHA-256 hash and a current or superseded status. Version releases so filed evidence is never edited, and start with an entity map showing who holds each record after acquisitions, carve-outs or shared systems.

Retrieval

Portfolio data room dataset

This page covers portfolio data room dataset evidence: an index and version convention.Illustrative

Key takeaways

  • Index evidence by stable ID and dataset version instead of relying on folder names.
  • Treat every released dataset version and every filed document as frozen; changes become new versions and new rows.
  • Record SHA-256 digests for data files and evidence documents so both sides can confirm what was reviewed.
  • Begin with an entity map, because rights follow the entity that created the records, not the platform storing them.
  • Hold back records from divested or unknown entities until counsel confirms which terms apply.

How do you build a dataset evidence folder for buyer diligence?

Build it around a single evidence index rather than a tree of folders. Every document gets a stable evidence ID, the entity and source system it covers, the dataset version it supports, a content hash and a status of current or superseded. A reviewer can then check any claim about the dataset against a specific, unaltered file, and see what changed between releases.

Portfolio and transaction settings make this harder than a single-company sale. Records may have been created by a company that was later acquired, shared through a transition services arrangement, merged into a platform company's systems, or partly divested. A folder that only collects documents cannot show which approval covers which records in which version, and that is the question a careful buyer asks first.

The convention on this page is meant for deal teams, portfolio operating partners and the data owners who do the work. It assumes the company is preparing a defined dataset for review, not a whole data room.

Where a data room checklist stops once the dataset keeps changing

A checklist tells a team what to collect; it does not tell a reviewer which version of the dataset each document describes. Datasets are re-extracted, corrected and narrowed during diligence, and evidence that was true for the first extract may be false for the third.

Software teams solved a version of this problem long ago. The Semantic Versioning specification says that once a versioned package is released, its contents must not be modified, and any modification must be released as a new version. Applying the same rule to datasets, and to the evidence folder that describes them, keeps a buyer from reviewing a document that quietly no longer matches the delivered files.

The second gap is integrity. NIST's Secure Hash Standard describes hash algorithms such as SHA-256 as a way to detect whether a message has changed since its digest was generated. Recording a SHA-256 digest for each dataset file and each evidence document lets either side confirm later that what was reviewed is what was delivered.

The evidence index: one row per piece of proof

The index is the asset of this page: a single table, kept as a spreadsheet or CSV at the root of the folder, with one row per document. A team can adopt the fields below as is.

  • Never edit a filed document; file the new one under a new ID and mark the old row superseded, so the history of what a buyer saw survives.
  • Write the claim in plain words; a row whose claim cannot be stated in one sentence usually needs splitting.
  • Keep personal data out of the folder; an export log or field list proves what a system holds without copying the records.
The evidence index: one row per piece of proof
FieldWhat it holdsExample (illustrative)Who maintains it
Evidence IDStable identifier that never changes or gets reusedEV-0042Folder owner
Evidence typeContract, consent or notice version, system export log, approval, transformation record, hash manifestCustomer terms, version 3Folder owner
Entity and systemLegal entity whose records it covers and the source systemAcquired subsidiary; its help desk instanceData owner
Period coveredDate range of records the document applies toRecords created in years 2 to 5 after acquisitionData owner
Supports claimThe dataset statement this document provesCustomers were notified of service-improvement useDeal counsel or data owner
Dataset versionsRelease numbers the document is valid for1.0.0 to 1.2.0Folder owner
Content hashSHA-256 digest of the file as filed64 hexadecimal charactersFolder owner
StatusCurrent, superseded by another ID, or withdrawn with a reasonSuperseded by EV-0057Folder owner
Approver and dateWho confirmed the row and whenGeneral counsel, with dateNamed approver

A version convention that ties every proof to one dataset release

Give each dataset release a three-part number and change the part that matches the kind of change. The rule borrowed from Semantic Versioning is simple: a released version is frozen, and the number tells a reviewer how much of the evidence they need to read again.

The mapping below matters most late in a process, when a buyer has already reviewed one release and the seller corrects or narrows it. A patch or minor release lets the buyer limit their second look to the rows that changed. A major release signals that the rights or entity picture moved, and that earlier conclusions should not be carried forward without a fresh read. Teams that skip numbering tend to send a "final_v3" file that nobody can match to the documents reviewed two weeks earlier.

  • Name files with the evidence ID and version, such as EV-0042_v1.1.0, and never reuse a name for different content.
  • Keep a manifest per release listing every dataset file with its SHA-256 digest, record count and the evidence IDs that support it.
  • Write a one-paragraph change log per release that names the rows added, superseded or withdrawn.
A version convention that ties every proof to one dataset release
Change to the datasetVersion changeEvidence to updateBuyer re-review
Corrected values or fixed redaction misses, same scope and fieldsPatch, such as 1.0.0 to 1.0.1Hash manifest and a correction noteSpot check of the changed files
More records from the same entities, systems and period rulesMinor, such as 1.0.1 to 1.1.0Hash manifest, volume note, export logNew records and their provenance rows
New field or record family within the same rights basisMinorField list, transformation record, privacy review noteThe added fields
Records removed because an entity or period left scopeMajor, such as 1.1.0 to 2.0.0Entity map, supersession rows, removal logFull review of rights rows
New entity, system or rights basis addedMajorNew contract and notice rows, entity map, approvalsFull review of rights and privacy rows

Which entity has authority over each record after the transaction or portfolio change?

Start every folder with an entity map, because rights follow the entity that created or received the records, not the platform that now stores them. A buyer reviewing a portfolio dataset will ask which company's terms applied to each record when it was created, and whether that company or a successor still holds the right to license it.

Draw the map before collecting documents. List every legal entity whose records could appear in the dataset, the dates each joined or left the portfolio, and the systems that held its records in each period. Each row of the evidence index then points to one entity, which keeps a contract signed by one company from being read as covering another company's customers.

Which entity has authority over each record after the transaction or portfolio change?
SituationWho usually holds the recordsEvidence the index needsTypical open question
Add-on acquisition merged into the platform's systemsThe acquiring entity, subject to the target's prior termsPurchase agreement data provisions, target's customer terms by dateDo pre-closing terms permit the proposed use?
Carve-out with a transition services arrangementSeller and buyer split by business lineSeparation schedule, system access terms, data split logWhich records belong to the divested business?
Shared system across portfolio companiesEach company for its own recordsTenant or field that identifies the entity, access controlsAre records cleanly separable by entity?
Duplicate customers across portfolio companiesEach company under its own termsDeduplication log naming which entity's copy was keptDid dedup move a record under a different contract?
Vendor-hosted platformThe company, subject to the vendor contractVendor agreement data clauses, export confirmationDoes the vendor agreement restrict downstream use?

Mapping the folder to datasheets, Croissant and provenance vocabularies

The folder is easier to review when its structure echoes documentation formats buyers already use. Datasheets for Datasets, proposed by Gebru and colleagues, groups questions into motivation, composition, collection process, preprocessing and labeling, uses, distribution and maintenance; those headings make sensible top-level sections for the index's claims.

MLCommons' Croissant format offers machine-readable fields for the same ideas. Its core specification lets each file object carry a sha256 checksum, and its responsible AI extension adds properties such as rai:dataCollection, rai:dataCollectionTimeFrame, rai:personalSensitiveInformation, rai:dataUseCases and rai:dataLimitations. Pointing each of those fields to evidence IDs links the metadata a buyer loads to the documents that justify it.

W3C's PROV-O vocabulary supplies the grammar for lineage: entities, activities and agents, with relations such as prov:wasDerivedFrom between a dataset version and the extract it came from. A team does not need to publish PROV-O to benefit; using its three nouns in the transformation records keeps every step answerable to who did what, to which file, and when.

Illustrative worked example: an add-on acquisition and a later carve-out

The scenario is invented to show the convention at work; the counts are illustrative. A platform company in a services portfolio prepares a dataset of support tickets from itself and from an add-on it acquired, then must remove records when part of the add-on is later sold.

  • Release 1.0.0: tickets from both entities, with the entity map, both companies' customer terms by date, export logs and a manifest of 12 files with SHA-256 digests.
  • Release 1.0.1: a reviewer finds names left in a free-text field in about 2 percent of tickets; the redaction is rerun, the manifest is reissued and a correction note is filed.
  • Release 1.1.0: six further months of tickets from the same systems are added under the same terms; new export logs and a volume note join the index.
  • Release 2.0.0: a business line of the add-on is sold; its tickets are removed, the separation schedule is filed, rights rows for that line are marked withdrawn and the buyer re-reviews the rights section in full.
  • Exception: the deduplication log shows some customers' surviving copies came from the divested line; those records are held back until counsel confirms which entity's terms apply.

Questions the folder cannot settle without counsel

The folder organizes evidence; it does not decide what the evidence means. Whether a purchase agreement transferred the right to license records, whether pre-closing notices cover a new use, and whether a separation schedule left any records with the seller are questions for qualified counsel working from the specific documents.

Regulators have shown why the link between evidence and a specific dataset version matters. In its 2021 settlement with the photo app developer Everalbum, the FTC required the company to delete models and algorithms it had developed using users' photos and videos, along with face embeddings derived from the photos of users who had not consented. A folder that can show exactly which records, under which notice, went into which release is what lets a team answer that kind of question quickly.

Pause the release and ask for specialist review when the index shows records whose creating entity is unknown, a notice version that does not cover the proposed use, a shared system that cannot separate entities, or any record from a divested business. Access to a system, a privacy permission, ownership of records and a contractual right to license them are four separate questions, and de-identification does not answer the last one.

Where this folder meets a SourceX licensing review

When the folder backs a dataset licensed through SourceX, its version numbers and manifests are what make the signed scope precise, because the company approves that scope and its terms, and signs, before any data is delivered. Large deliveries remain in the seller's own storage or ship on encrypted drives, so a manifest with file hashes is also how both sides confirm a delivery arrived whole. Ahead of all this sits a brief assessment of fit that the company can take without uploading data.

Frequently asked questions

What is a dataset evidence folder?

It is the set of documents that proves where a dataset's records came from, which terms governed them, how they were transformed and what was delivered. Organized around an index with stable IDs and version numbers, it lets a buyer check each claim against a specific file.

How is it different from a sell-side data room?

A data room covers the whole company for a transaction. A dataset evidence folder covers one defined dataset and follows it through releases, so each document is tied to the versions it supports.

Why hash evidence documents as well as data files?

A digest recorded when a document is filed lets anyone confirm later that it has not been altered. That matters when a reviewer relies on a specific version of a contract, notice or export log.

Who should own the evidence index?

One named folder owner keeps the index consistent, while data owners and counsel approve the rows in their areas. Splitting ownership across deal team members without one owner is how superseded documents end up marked current.

Does this replace legal diligence on the transaction?

No. The folder makes evidence findable and versioned; counsel still decides what the transaction documents, notices and contracts allow for each entity and period.

Sources

  • Semantic Versioning requires that a released version's contents never be modified and defines when to increment major, minor and patch numbers. Source
  • NIST's Secure Hash Standard (FIPS 180-4) specifies SHA-256 and describes hash digests as a way to detect whether a message has changed. Source
  • Datasheets for Datasets organizes dataset documentation into motivation, composition, collection process, preprocessing and labeling, uses, distribution and maintenance. Source
  • The Croissant specification lets each FileObject carry a sha256 checksum of its content. Source
  • The Croissant RAI extension defines properties for data collection, collection time frame, personal and sensitive information, use cases and limitations. Source
  • PROV-O defines Entity, Activity and Agent as starting-point classes and wasDerivedFrom for derivation between entities. Source
  • The FTC's Everalbum settlement required deletion of models and algorithms developed with users' photos and videos, and of face embeddings derived from users who had not consented. Source

See if your company qualifies

A short company assessment. No data uploads are needed.

See if your company qualifies