Skip to content

Rights and contracts

Dataset provenance attestation limits: what a supplier's signature proves

By SourceX Editorial · Reviewed by Noah Loul ·

9 sections

Short answer

A supplier attestation proves who made a claim about a dataset, when, with what authority and on what terms of accountability. It does not prove that the records came from the named system, that permission covers the intended use, or that every record was transformed as described. Those facts need record-level evidence: a source register, transformation versions, permission references, release IDs and checksums.

Provenance trail

Dataset provenance

This page covers dataset provenance attestation limits: what a supplier's signature proves.Illustrative

Key takeaways

  • An attestation proves who made a claim, when and with what accountability; it does not prove the underlying facts.
  • Origin, completeness, transformation and licensing permission each need their own record-level evidence.
  • Tie every claim to a chain of IDs: record, source system, extraction run, transformation version, permission reference and release.
  • Rank supporting evidence by how hard it would be to fake, and disclose where the attestation rests on weak support.
  • Narrowing a claim to what the records show is better than signing a broad statement a buyer later disproves.

What can a supplier attestation establish about dataset rights and provenance?

A supplier attestation establishes accountability, not facts. It records that a named person, acting for a named company, stated specific things about a dataset on a specific date and accepted the consequences of being wrong. That is valuable, but it is a promise about evidence rather than the evidence itself.

The practical limit is simple: an attestation is only as strong as the records behind each sentence in it. A sentence such as "all records were collected from our service management system between 2019 and 2025" can be true, half true or unknowable, and the signature reads the same in all three cases. A buyer, a reviewer or your own counsel can only tell the difference by tracing the claim to the source register, the extraction run, the transformation version, the permission reference and the release that delivered it.

This page is written for the data owner who is asked to sign. It sets out what the signature proves, what it cannot prove, and an attestation evidence checklist that lets you back each claim before you sign it. It does not replace a contract review; the wording of any attestation you give should be read by counsel against the license it sits in.

What does a signed attestation actually prove?

A signed attestation proves three things: identity, time and commitment. It shows who made the statement, when it was made, and that the signer accepted a defined form of responsibility for it. In U.S. federal practice, for example, 28 U.S.C. 1746 lets a written, dated statement signed under penalty of perjury stand in for a sworn affidavit where federal law or rules require one, and the statute supplies the model wording. Commercial data attestations usually rest on contract rather than perjury law, but the logic is the same: the document converts a claim into an accountable claim.

Self-attestation is a recognized tool in procurement. CISA's Secure Software Development Attestation Form asks a software producer to attest to a set of secure development practices, and it shows the limits of the approach in its own design: a producer may submit a third-party assessment instead, and agencies can ask for additional artifacts such as a software bill of materials. The attestation alone is treated as a starting point that artifacts and independent review can strengthen.

For a dataset, that means a signature can do real work. It allocates responsibility, it fixes the claim in time so later changes can be measured against it, and it signals that someone with authority looked at the question. What it cannot do is substitute for the trail of records that would let a stranger reach the same conclusion.

What does a supplier attestation not prove?

An attestation does not prove any fact that the signer could not personally verify, and most dataset facts fall in that category. The person who signs is usually an executive or data owner, not the engineer who ran the export or the administrator who configured the source system ten years ago.

The gaps listed on this page do not make an attestation worthless. They define what it should be paired with. The useful question for each sentence is not whether it is true, but what a stranger would need to see to agree that it is true, and whether that material exists today.

  • Origin. A statement that records came from a named system does not show which export, which instance or which date range produced the delivered files. Only an extraction log, a source register entry and file checksums tie the delivery to the system.
  • Completeness. A statement that the dataset covers a period says nothing about gaps from migrations, archived records, purged history or failed jobs.
  • Transformation. A statement that personal data were removed does not show that the same transformation version ran on every record, or that free-text fields and attachments were included in its scope.
  • Permission for this use. Having the right to access, store or process records is different from having a contractual right to license them for AI training. Privacy permission, ownership and licensing rights are separate questions, and de-identification alone does not create a right to license.
  • Third-party content. Records often contain material written or owned by customers, vendors or contractors. A general statement of ownership rarely addresses each of those contributors.
  • Continued truth. An attestation is true as of its date. A permission that is later withdrawn, a source system that is replaced or a correction to old records can make it stale without anyone noticing.

The attestation evidence checklist

Use this checklist before signing, one row per claim the attestation will make. If you cannot point to the evidence in the second column, rewrite the claim to what you can support or attach the gap as a stated limit. Version 1.0, dated 2026-10-18.

Work through it with three people in the room: the person who will sign, the person who ran or owns the extraction, and the person who holds the contracts and policies that grant rights over the records. Each row gets an owner, a pointer to the evidence and a short note on its strength. Rows without evidence are not failures; they are the limits section of the attestation, written before a buyer has to discover them.

Keep the completed checklist with the release it describes. When a buyer asks a follow-up question months later, or when a later release changes the scope, the checklist shows exactly which evidence supported which sentence on the day it was signed.

The attestation evidence checklist
Attested claimEvidence that backs itWhere it usually livesGap if missing
Records came from the named source systemSource register entry with system name, instance, owner and export method; extraction run IDData team runbook, export job logsOrigin rests on memory; narrow the claim to systems you can log
Records cover the stated date rangeRow counts by month, earliest and latest timestamps, known gaps listedProfiling output from the extraction runSilent gaps from migrations or archiving
Files delivered match the files reviewedChecksums per file recorded at release, release IDRelease manifestNo way to show a file was not swapped or altered
Personal data were removed or maskedTransformation name and version, field list in scope, sample test resultsPipeline repository, test reportsClaim covers some fields or some records only
The company may license these records for AI usePermission reference per source: contract clause, policy or consent record, and who reviewed itLegal file, contracts registerAccess or privacy permission mistaken for licensing rights
Third-party content is excluded or clearedRule that identifies third-party fields and attachments, and the count removedPipeline rules, exception logCustomer or vendor material inside free text
The signer has authority to bind the companyBoard or officer delegation, title and signature dateCorporate recordsAttestation may be challenged as unauthorized
Known limits are disclosedLimits section listing exclusions, gaps and uncertain sourcesAttestation annexBuyer relies on a cleaner picture than exists
Changes after signing will be notifiedChange notice procedure with owner and trigger listData governance policyStale attestation remains in use

How do you tie each claim to the exact record and release?

Tie claims to records with a chain of identifiers that a reviewer can follow in either direction: record ID to source system, extraction run, transformation version, permission reference and release ID. If any link is missing, the claim cannot be traced for that record, and the attestation should say so.

Existing standards give you a vocabulary for that chain. W3C PROV-O models provenance with three core classes, Entity, Activity and Agent, and relations such as wasGeneratedBy, wasDerivedFrom and wasAttributedTo. In plain terms, each delivered file or record is an entity, each export or transformation run is an activity, and each person, team or software tool responsible is an agent. The Croissant metadata format for machine learning datasets builds its provenance representation on PROV-O and recommends recording checksums for files in versioned datasets, which is how a buyer later confirms that the files in hand are the ones described.

Version 1.0.0 of the Data & Trust Alliance Data Provenance Standards groups 22 metadata fields into three standards, Source, Provenance and Use, with later updates planned through an open community process. Mapping your attestation sentences to fields in a structure like that keeps the claims and the metadata in step: if the attestation says something the metadata does not record, one of them is wrong.

Regulators are moving the same way. For high-risk AI systems, Article 10(2)(b) of the EU AI Act requires data governance practices to cover data collection processes and the origin of data, and for personal data the original purpose of collection. The NIST AI RMF Playbook asks how an organization has documented data provenance, including sources, origins, transformations, augmentations, labels, dependencies, constraints and metadata. A buyer subject to either framework will want records, not only a signature.

Datasheets for Datasets, the documentation template proposed by Gebru and coauthors, frames the same chain from the consumer's side: its questions on collection process, preprocessing and maintenance are the ones a buyer will ask when an attestation is not backed by records.

How strong is each kind of supporting evidence?

Rank evidence by how hard it would be to produce if the claim were false. A statement in an email is weak; a log written automatically by the system at the time of export is strong; a sample a second person re-performed from the source is stronger still. Record the strength next to each claim so a reviewer can see where the attestation leans on weak support.

How strong is each kind of supporting evidence?
StrengthEvidence typeWhat it showsLimit
1 WeakUnsupported statement by the signerWhat the company believesNo way for a third party to check it
2 ModerateDocuments prepared for the attestation, such as a source register written afterwardA considered, reviewable accountCan repeat the same error as the statement
3 StrongSystem-generated records made at the time: export logs, job IDs, checksums, audit trailsThat an event happened as recordedShows what ran, not whether the scope was right
4 StrongerSample re-performed from the source by someone other than the preparerThat records trace back as claimed for the sampleCovers the sample, not every record
5 StrongestIndependent review against a written method, with findings reportedAn outside view of the method and the evidenceCostly; still limited to what the reviewer examined

Worked example: a maintenance ticket dataset with one exception

This example is illustrative; the company and figures are hypothetical. A property management company prepares six years of maintenance tickets from its help desk system for licensing. The draft attestation, written by the operations director, says the records came from that system, cover 2019 to 2025, contain no tenant personal data and may be licensed for AI training.

Working through the checklist changes three sentences. The export logs show the system was adopted in mid-2020; earlier tickets were imported from spreadsheets, so the origin claim is split into two sources with separate register entries. Profiling shows a four-month gap after the import, which is listed as a known limit rather than left for the buyer to find.

The masking claim survives with a narrower scope. The masking pipeline ran on structured contact fields and ticket comments, version-tagged and tested on a sample, but photo attachments uploaded by tenants and outside contractors were never in scope. The exception is handled by excluding attachments from the release and stating the exclusion in the attestation.

The licensing claim is the one the company cannot fully support on its own. Its contracts with outside contractors say nothing about reuse of their notes. Counsel narrows the dataset to tickets and comments written by employees and attaches the permission reference for that scope. The final attestation is shorter than the draft, and every remaining sentence points to a record.

The release manifest then records the checksums of the delivered files, the masking pipeline version and the permission reference, all under one release ID. If the company later corrects ticket categories or removes a building it sold, the manifest shows which release the attestation described and which change needs a notice. The exception, the attachments, is visible in three places: the limits section, the pipeline rules and the manifest's file list.

When should a signer refuse, narrow or escalate an attestation?

Stop and narrow the claim whenever the evidence column of the checklist is empty for a sentence you are being asked to sign. Narrowing is normal and usually better for both sides than a broad claim that a buyer later finds is unsupported.

  • Stop if no one can identify the system instance or export that produced the files. Re-run the extraction with logging before attesting.
  • Narrow the scope when part of the data has a different origin, such as an import from a predecessor system or an acquired company.
  • Narrow the scope when free text or attachments contain third-party content you cannot identify reliably.
  • Request legal review when the licensing right depends on contract terms, customer agreements, employee policies or consent records.
  • Request privacy review when the claim concerns de-identification, sensitive categories or records about individuals in more than one jurisdiction.
  • Request a technical review when transformation claims depend on code that has changed between runs.
  • Do not sign an attestation that promises future facts, such as the absence of all errors, unless the agreement defines how errors are corrected and notified.

Where the attestation fits in a licensing deal

In a licensing deal the attestation comes last, after the rights review it summarizes. On SourceX, the rights review is one step in a single process that also covers qualification, contracting, preparation, delivery and payment for selected business records, and the company's signature on scope and terms comes before any delivery.

Qualification starts with a brief assessment of the company that needs no uploaded records. Starting the evidence checklist at that stage, rather than when an attestation lands on the signer's desk, keeps the final statement narrow enough to stand behind.

Frequently asked questions

Is a supplier attestation the same as a warranty?

Not necessarily. An attestation is a statement of fact by the signer, while a warranty is a contractual promise with remedies if it proves untrue; a license may contain both. Counsel should read how the two interact in your agreement.

Can a buyer rely on an attestation without seeing evidence?

A buyer can, but it takes on the risk that the claim is wrong and has to rely on contract remedies afterward. Many buyers ask for supporting artifacts or test a sample of records against the source. Offering the evidence checklist up front usually shortens that process.

Who should sign a dataset provenance attestation?

Someone with authority to bind the company, typically an officer or a delegated data owner. That person should sign only after the people who ran the extraction and transformations have confirmed the evidence for each claim in writing.

Does de-identification make it safe to attest that the data can be licensed?

No. De-identification addresses privacy risk; it does not create ownership or a contractual right to license. The permission to license has to come from contracts, policies or other rights, reviewed by counsel.

How long does an attestation stay valid?

It describes the dataset as of its date. Corrections, withdrawn permissions, system changes or new releases can make it stale, so the agreement should define which changes trigger a notice and an updated attestation.

Sources

  • Under 28 U.S.C. 1746, a written, dated statement signed under penalty of perjury can stand in for a sworn affidavit where federal law or rules require one. Source
  • CISA's Secure Software Development Attestation Form is a producer self-attestation; a third-party assessment may be submitted instead, and agencies can request additional artifacts such as an SBOM. Source
  • W3C PROV-O is built on three core classes, Entity, Activity and Agent, with relations such as wasGeneratedBy, wasDerivedFrom and wasAttributedTo. Source
  • Croissant 1.1 builds its provenance representation on W3C PROV-O, recommends file checksums in versioned datasets, and defines PATCH, MINOR and MAJOR version increments for dataset content. Source
  • Version 1.0.0 of the Data & Trust Alliance Data Provenance Standards defines 22 metadata fields grouped into three standards: Source, Provenance and Use. Source
  • Article 10(2)(b) of the EU AI Act requires data governance practices for high-risk AI training, validation and testing data to cover data collection processes and the origin of data, and for personal data the original purpose of collection. Source
  • The NIST AI RMF Playbook asks how an organization has documented an AI system's data provenance, including sources, origins, transformations, augmentations, labels, dependencies, constraints and metadata. Source
  • Datasheets for Datasets proposes documentation covering motivation, composition, collection process, preprocessing, distribution and maintenance, including how updates will be communicated to dataset consumers. Source

See if your company qualifies

A short company assessment. No data uploads are needed.

See if your company qualifies