Skip to content

Privacy and preparation

Dataset procurement, manifest first: 8 questions before raw records move

By SourceX Editorial · Reviewed by Noah Loul ·

8 sections

Short answer

Manifest first means a buyer asks for a records-free description of the dataset before any raw record moves: files, counts, date ranges, field definitions, label method, declared personal information, per-source rights basis, file hashes and maintenance plan. In this article's eight-question map, four questions settle on the manifest and three narrow to a small check; only task fit always needs records. Request records for open questions only.

Redaction

Dataset procurement

This page covers dataset procurement, manifest first: 8 questions before raw records move.Illustrative

Key takeaways

  • A manifest describes records without containing them, so most early fit questions can be answered before any personal or confidential data moves.
  • In the eight-question map, four questions settle on the manifest, three narrow to a targeted check, and only task fit always needs records.
  • Data minimization rules point the same way: GDPR limits personal data to what is necessary for the purpose, and HIPAA's minimum necessary standard covers requests as well as disclosures.
  • Ask for fields that published standards already define, such as per-file hashes and byte counts, per-source licenses and declared personal or sensitive information.
  • A manifest is a claim: pair every line with the check that will test it at sample or delivery, and gate sample requests on rights and personal information.

What does manifest first mean in dataset procurement?

It means the buyer's first request is for a description of the dataset, not the dataset. A manifest lists what the package contains, file by file, with counts, date ranges, field definitions, how labels were made, what personal information was present and removed, the rights basis for each source, file hashes and the maintenance plan. It contains no records.

The rule that follows is short: raw records move only to answer a fit question the manifest cannot. Manifest first does not replace a sample. It decides whether a sample is worth requesting, what it must contain and how small it can be.

This is a buyer-side discipline. A supplier describing its holdings early is a separate practice; the question here is what an AI team should ask a vendor or data owner for, in what order, and what each answer can and cannot prove.

Why asking for raw records first is the costly mistake

Because raw records carry obligations from the moment they arrive, and most early questions do not need them. A buyer that receives a full export to answer a scope question now holds personal and confidential data it may have no basis to keep, and the supplier has disclosed that data before any terms were agreed.

The two regimes that most often govern business records point the same way. Article 5(1)(c) of the GDPR requires personal data to be adequate, relevant and limited to what is necessary for the purposes of processing, and Article 25(2) applies data protection by default to the amount of personal data collected, the extent of processing, the storage period and accessibility. For U.S. health records, HHS guidance on HIPAA's minimum necessary standard says a covered entity must make reasonable efforts to use, disclose and request only the minimum protected health information needed for the purpose, and may not use, disclose or request an entire medical record without specific justification.

A supplier bound by either regime has to ask why your request needs what it asks for. A manifest-first request answers that before it is asked, and it keeps the first exchange out of the supplier's heaviest approval path because nothing sensitive is released. This is general information; counsel should confirm what applies to a particular dataset.

Eight fit questions and the manifest line that settles each

Four of the eight questions a buyer asks before purchase can be settled from a good manifest, three can be narrowed to a small targeted check, and one, whether the data improves your task, always needs records. The map is a scorecard: mark each question settled, narrowed or open for every vendor.

Use it in order. A question settled on the manifest never justifies a records request; a narrowed question justifies only the slice or document needed to close it; the open question justifies a sample sized for your acceptance test. Each row names the manifest line that answers the question and the next step if it stays open:

Eight fit questions and the manifest line that settles each
Fit questionFrom the manifestManifest line that answers itIf still open, escalate to
Do the systems, years, record types and per-slice volumes match the need?SettledFile list with source system, date range and record count per file and per sliceNothing: a mismatch here ends the conversation
Do the required fields exist and mean what the task assumes?NarrowedData dictionary with definitions, types, null rates and changes in meaning over timeA few dozen de-identified records limited to those fields
Are outcomes or labels present, and how were they produced?NarrowedLabel definitions, who or what assigned them, and any agreement checksA small de-identified sample scored against your rubric
What personal or sensitive information was present, and what was removed?Settled as a declarationDeclared personal and sensitive information, removal method and how removal was checkedA residual personal data audit on the evaluation sample
Which rights basis covers each source?NarrowedPer-source provenance with its license or rights basisContracts and notices reviewed under NDA, not records
Will the delivery be provably what was described?SettledPer-file byte counts and content hashes, with a version stringNothing until delivery, when the hashes are recomputed
Will the dataset change after purchase?SettledMaintenance plan: planned updates, support for older versions, notice of changesNothing: write the answer into the license schedule
Does the data improve the task?OpenNone; a manifest cannot answer thisAn evaluation sample under an evaluation-only license, sized for your acceptance test
A dot matrix of eight fit questions against three evidence stages, manifest, targeted check and evaluation sample, showing four questions settled on the manifest, three narrowed and closed by a targeted check, and only task lift needing a sample.
The manifest-first question map. Editorial framework; statuses describe a complete manifest, not any vendor.

What a buyer-grade manifest should contain

Ask for fields that published dataset standards already define, so the vendor cannot answer with a brochure. Two open specifications cover most of the list.

The Frictionless Data Resource specification gives each file a size in bytes and a hash, and lets each resource carry its own sources and licenses instead of inheriting one statement for the whole package. Croissant, the MLCommons format for machine learning datasets, gives file objects a sha256 checksum and strongly recommends recording checksums for versioned datasets; its responsible AI extension adds a field describing personal and sensitive information in the data, plus fields for the collection time frame and the maintenance plan.

The Datasheets for Datasets framework supplies the maintenance questions a buyer should not skip: whether the dataset will be updated and whether older versions stay supported. NIST's Generative AI Profile (NIST AI 600-1) suggests documenting the origin of training data, including licenses and data privacy, which is the provenance column of the same manifest. A fictional, finished example of a package summary and field guide is on the SourceX sample manifest page. A working request covers eight blocks:

  • Package: name, version string, cutoff date, preparer and approver.
  • Files: path, format, byte count and content hash for every file.
  • Records: counts per file and per slice that matters to your task, such as year, category, language or outcome.
  • Fields: a data dictionary with definitions, types, null rates and any change in meaning across the date range.
  • Labels: what each label means, who or what assigned it, and how agreement was checked.
  • Personal and sensitive information: what was present, what was removed or transformed, and how removal was checked.
  • Provenance and rights: each upstream source with its license or rights basis and any restriction on AI use.
  • Maintenance: planned updates, whether older versions stay available, and how changes will be notified.

When the manifest is not enough: escalate one question at a time

Escalate only for the questions still open, and only by the smallest step that can close them. A narrowed question about field meaning needs a few dozen records restricted to those fields, not the full export; a rights question needs contracts and notices under NDA, not records at all.

The open question, whether the data improves your task, needs a real sample. Size it from your acceptance test, draw it at random across the date range, keep it de-identified, and hold it under an evaluation-only license that limits use and sets a deletion date. Where the supplier cannot release even that, restricted routes such as view-only access, redacted samples or remote queries exist, each with limits on what it can prove.

Illustrative: a team buying support-ticket histories for an agent receives a manifest with three years of tickets, a data dictionary and per-file hashes, but labels described only as resolution codes. Scope, delivery match and future changes settle on the manifest. Field meaning and labels narrow, so the team asks for 40 de-identified tickets limited to the status, resolution code and thread fields, to confirm that a resolution code means solved rather than merely closed. Rights narrows to the customer terms and privacy notice, reviewed under NDA. Only then does the team request an evaluation sample sized for its acceptance test, under an evaluation-only license.

Keep a log of what moved and why. If a later audit asks why your team held a set of records, the answer should be one open question per transfer, each tied to a manifest line that could not settle it.

Can a polished manifest mislead a buyer?

Yes. A manifest is the vendor's claim about its data, and its polish says little about the records unless every line is tied to a later check.

At a glance

Can a polished manifest mislead a buyer

License fields are the clearest case. According to the abstract of the Data Provenance Initiative's paper, which audited more than 1,800 text datasets, widely used hosting sites left licenses unstated in more than 70 percent of cases and misreported them in more than 50 percent. A tidy rights column can still be wrong, which is why the map sends rights questions to documents rather than accepting the declaration.

The same holds for the evidence vendors offer instead of a manifest. Catalog size measures inventory, not fit. A security report covers the vendor's systems, not what sits inside a dataset. A case study describes another buyer's task. None of them settles a row in the map, and none replaces acceptance evidence on your own sample. Pair each manifest line with the check that will test it:

  • Counts and date ranges: recounted on delivery.
  • Hashes and byte sizes: recomputed for every delivered file.
  • Field definitions and labels: tested on the evaluation sample.
  • Declared personal information and removals: residual audit on the sample and again at delivery.
  • Rights basis: matched to the contracts and notices reviewed under NDA.
  • Maintenance plan: written into the license schedule with a notice obligation.

Scoring vendors before any record moves

Score every vendor on the same map before requesting a sample, and let two rows act as gates. A vendor that cannot describe personal and sensitive information or the rights basis per source is not ready for a sample request, however complete the rest of its manifest looks.

Illustrative scoring: rate the seven rows a manifest can address at 2 for a specific, checkable answer, 1 for a vague one and 0 for none, for a maximum of 14. Treat a 0 on personal information or rights as a stop, and request an evaluation sample only from vendors at 10 or more with no stop. The thresholds are editorial; tune them to the sensitivity of the records and the cost of a sample review.

Record the outcome in a short decision memo so the next reviewer can see why records moved. It needs six fields:

  • Vendor, dataset name, manifest version and the date it was received.
  • Status of each row: settled, narrowed or open, with the manifest line cited.
  • Gate results for personal information and rights basis.
  • Records requested, the open question each request answers, and the slice limits.
  • The license that covers the sample and its deletion date.
  • Decision: proceed to a sample, ask for a better manifest, or stop, with the owner and date.

A manifest-first request to SourceX

Start with the requirement, not a records request. SourceX holds no stock and sources each dataset from U.S. businesses once a buyer has asked; the request-data form, which already takes language, format, scope, volume, timing and licensing needs, is where the manifest lines you need and your acceptance test belong.

The same order applies on the supply side. A company starts with SourceX through a short assessment that needs no data uploads, and it approves scope and terms and signs its agreement before a single record is delivered to SourceX. A buyer requirement written as manifest questions first and sample criteria second fits that sequence.

Frequently asked questions

What is a dataset manifest?

A dataset manifest is a records-free description of a data package: files, formats, byte counts and hashes, record counts and date ranges, field definitions, label methods, declared personal information, per-source rights basis and maintenance plan. It lets a buyer judge scope and fit before any record moves.

Is a data dictionary the same as a manifest?

No. A data dictionary defines the fields; a manifest also covers files, counts, date ranges, provenance, rights, removals, hashes and versioning. The dictionary is one block of a complete manifest.

Can a manifest itself contain personal data?

It should not. Use field names, definitions and counts instead of example values copied from real records, and avoid counts so small that they point to a person. Where an example value helps, make it fictional.

How many records should a buyer request after reading the manifest?

Only enough to close the questions still open. Field meaning can often be checked on a few dozen de-identified records limited to the relevant fields, while task fit needs a random sample sized from your acceptance test and held under an evaluation-only license.

Does manifest first slow procurement down?

Usually not. Vendors that cannot describe their data drop out before anyone reviews records, a records-free request asks less of the supplier's approval process than an export, and the sample review that follows is smaller because it targets only the open questions.

Sources

  • GDPR Article 5(1)(c) limits personal data to what is adequate, relevant and necessary for the purposes of processing, and Article 25(2) applies data protection by default to the amount collected, extent of processing, storage period and accessibility. Source
  • HHS guidance on HIPAA's minimum necessary standard requires covered entities to make reasonable efforts to use, disclose and request only the minimum protected health information needed, and to justify any request for an entire medical record. Source
  • The Frictionless Data Resource specification defines per-file bytes and hash properties and lets each resource list its own sources and licenses. Source
  • The Croissant 1.1 specification gives file objects a sha256 checksum and strongly recommends recording checksums for versioned datasets. Source
  • The Croissant responsible AI extension defines a field describing personal and sensitive information in a dataset, along with collection time frame and maintenance plan fields. Source
  • Datasheets for Datasets includes maintenance questions about planned dataset updates and whether older versions remain supported. Source
  • NIST AI 600-1 suggests documenting the origin of training data, including licenses and data privacy. Source
  • The Data Provenance Initiative audit of more than 1,800 text datasets reports license omission above 70 percent and error rates above 50 percent on widely used dataset hosting sites. Source

Tell us what data you need

Share scope, volume, format, timing and licensing needs. Datasets are sourced on request.

See if your company qualifies