Deal economics
Dataset manifest investor diligence: a three-way match for data deals
By SourceX Editorial · Reviewed by Noah Loul ·
10 sections
Short answer
Dataset manifest investor diligence treats the manifest as the one document that states what a data company actually delivered. Run a three-way match: signed scope against manifest lines against acceptance records and invoices, sampled in both directions. Each line should carry a rights reference, a count with its unit, exclusions, a version and checksums. Lines that fail to match show revenue resting on promises the data has not yet met.
Value drivers
Dataset manifest investor
Key takeaways
- A manifest is the only record that says line by line what was delivered; scope documents state intent and invoices state money.
- The three-way match pairs signed scope, manifest lines and acceptance-plus-invoice records, sampled backward from invoices and forward from manifests.
- Each manifest line should carry a rights reference, a count with its unit, unique counts, exclusions, a pipeline version and file checksums.
- Major versions before acceptance signal rework; Croissant's guidance reserves a major version for edited, removed or reshuffled data.
- A clean manifest evidences fulfillment, not buyer value or buyer diversity; those need acceptance tests and demand evidence.
Why is a dataset manifest an investment document?
Because it is the only record that states, line by line, what a data company actually delivered, and every promise in the business has to reconcile to it. The license or deal confirmation says what may be delivered and on what terms; the invoice says what was billed; the manifest says which records, from which system and period, in which version, left the company. When the three disagree, the manifest is where the disagreement can be located.
A useful manifest lists, for each delivery, the record families and source systems, the date range, the record count with the unit it counts, the fields, the exclusions applied with their counts, the version of the preparation pipeline and a checksum for each file. Machine-readable dataset standards already carry several of these: MLCommons' Croissant format lets each file object declare a sha256 checksum, and lets a record set name a key, one or more fields whose values are unique across its records.
Buyers increasingly need these facts for their own obligations. Under the EU AI Act, providers of general-purpose AI models must publish a sufficiently detailed summary of the content used for training, following a template from the AI Office. California's AB 2013 requires generative AI developers to post training-data documentation, including a high-level summary of the datasets used in development, on or before January 1, 2026. A supplier whose manifests answer those questions shortens the buyer's work; one that cannot answer them adds to it.
How does the three-way match work for a data deal?
It pairs three records for each delivery, the signed scope, the manifest and the acceptance-plus-invoice record, and flags any line that appears in one but not the others. Accounts payable teams apply the same logic when they match a purchase order, a receiving report and a supplier invoice before paying; here the investor applies it to the seller's revenue.
Run it in both directions. Start from a sample of invoices and trace each billed line to manifest lines and then to signed scope, which tests whether revenue rests on delivered, permitted data. Then start from a sample of manifest lines and trace forward to acceptance and billing, which tests whether delivered data was accepted, rejected, reworked or handed over free as a sample or replacement. Mismatches are not automatically bad news; unexplained mismatches are.
Illustrative: an investor samples 20 invoice lines and 20 manifest lines across three buyers. Seventeen invoice lines trace cleanly, two point to a manifest version that was later replaced, and one has no manifest line behind it. Going forward, three manifest lines were never billed: two were free samples and one replaced a corrupted file. The invoice line with no delivery behind it is the finding to chase; the others have explanations the company can document.
The table lists the checks and what a failed match can point to:
| Match | Question it answers | Mismatch to look for | What it may mean |
|---|---|---|---|
| Scope to manifest | Was everything delivered inside the signed scope? | Record families, systems or years in the manifest that the scope never names | Delivery outran the grant, or the scope file is out of date |
| Manifest to acceptance | Did the buyer accept what was delivered? | Lines with no acceptance record, or accepted only in a later version | Rework, an open dispute or acceptance by elapsed time |
| Acceptance to invoice | Is billing tied to accepted lines? | Invoiced amounts with no accepted manifest line behind them | Revenue booked on a promise rather than a delivery |
| Invoice to scope | Do billed terms match the signed terms? | Volumes, periods or exclusivity on the invoice that differ from the scope | Side terms or amendments missing from the file |
| Version chain | Which version did the buyer finally accept? | Several major versions before acceptance | The quality promise was missed on first delivery |
Which manifest fields should an investor request?
Request the fields that tie each line to a promise, plus the companion evidence that lets you test each field instead of trusting it. This is the dataset diligence evidence request list for a manifest review:
| Manifest field | Promise it backs | Companion evidence | Test to run |
|---|---|---|---|
| Record family and source system | Scope | Signed scope or deal confirmation naming systems and families | Every family in the manifest appears in the scope |
| Date range and record count with its unit | Volume | The query or export log that produced the count | Recount one period; confirm whether the unit is rows, cases or messages |
| Unique entities or cases | Coverage | Key definition and duplicate report | Unique count holds near the claim after deduplication |
| Rights reference | Rights | The authority document each reference resolves to | Sampled lines reach a signer whose grant covers the buyer's use |
| Exclusions with counts | Privacy and obligations | Exclusion rules and the log of removed records | Source total minus removed equals delivered |
| Fields with missing-value rates | Quality | Profiling report on the delivered files | Rates on your own redraw match the manifest |
| Pipeline and de-identification version | Preparation | Change log for each version | The accepted version is the invoiced version |
| File checksums and delivery batch | Fulfillment | Transfer log and the buyer's receipt | Hashes of received files match the manifest |
How does a manifest connect rights to revenue?
Through a rights reference on each line, pointing to the document that authorized that record family for that use. Without it, an investor can confirm that data moved and that money moved, but not that the company was allowed to move it.
License metadata is fragile even in public data. Across the more than 1,800 text datasets it examined, the Data Provenance Initiative found widely used hosting sites dropping the license more than 70 percent of the time and recording it incorrectly more than 50 percent of the time. Business records licensed directly from companies avoid part of that problem because a named supplier signs, but the reference still has to survive packaging, splitting and redelivery, which is where line-level identifiers earn their cost.
Test it by sampling. Draw delivered lines at random, follow each rights reference to its authority document and check three things: the signer could grant that record family, the permitted use covers the buyer's use, and every exclusion the grant required shows up in the manifest's exclusion counts.
What should quality fields show beyond a record count?
They should show what was counted, how much is missing and what the count becomes after duplicates, because a count without its unit and denominator is a volume claim, not a quality claim. Illustrative: a manifest line reading 1.2 million records could mean 1.2 million messages, 90,000 tickets or 30,000 customers, and only one of those matches the buyer's unit of work.
Documentation frameworks already ask for these facts. Datasheets for Datasets proposes that each dataset come with a datasheet covering its motivation, composition, collection process and recommended uses. Croissant's responsible AI extension adds properties such as rai:dataCollectionMissingData, a description of missing data, and rai:dataBiases, a description of biases in the dataset. A manifest need not repeat a datasheet, but its quality fields should point to one and agree with it.
For an investor, the telling number is the gap between the manifest's claims and a redraw. Ask the company to regenerate counts and missing-value rates on a sample you choose; a gap nobody can explain says more about the operation than any single figure in the data room.
What do version changes reveal about fulfillment?
They reveal rework. Croissant's versioning guidance recommends semantic versions: a patch means the same data serialized or packaged differently, a minor version adds data while existing data stays retrievable, and a major version means existing data was edited, removed or shuffled across splits. A delivery that needed several major versions before acceptance did not meet its promise the first time.
Checksums make the chain auditable. Croissant strongly recommends recording checksums for all file objects in versioned datasets, so a recipient can confirm that the files it holds are the ones declared. If the manifest's hashes match the buyer's receipt for the accepted version, fulfillment is evidenced; if only the latest version carries hashes, earlier deliveries cannot be reconstructed.
Ask for the version chain of a few deliveries: each version's date, what changed and why, and which version the buyer accepted and was invoiced for. Count major versions after first delivery as a rework signal across the book, and compare that count with the company's own account of acceptance.
Does a clean manifest prove the business works?
No. A complete, matching manifest proves the company delivered what it said it delivered; it does not prove the data improved anything for the buyer or that the buyer will return. A signed agreement and a polished manifest can coexist with a buyer that tested the data, found no lift and let the license lapse.
Two checks keep the manifest in its place. First, look for acceptance against a written buyer test rather than acceptance by elapsed time. Second, look at demand concentration, because a perfect manifest from one buyer is still one buyer. Reddit's quarterly report for the period ended March 31, 2026 states that "substantially all of the contract value associated with our licensing revenue is derived from two of our partners," a reminder that even a large data licensor can depend on very few buyers.
Manifests can also be written after the fact. Prefer manifests generated by the delivery pipeline at packaging time, with timestamps and hashes, over spreadsheets assembled for the data room.
What could make manifest-based diligence mislead you?
The method trusts documents, and documents can be incomplete, retrofitted or beside the point. Name these risks before relying on a clean match:
- Retrofitted manifests. A spreadsheet built for diligence can match the invoices perfectly and describe nothing that was transferred; insist on pipeline-generated manifests and buyer receipts.
- Right files, wrong content. Hashes prove files are unchanged, not that their contents match the scope, so sample the records as well as the metadata.
- References that resolve to the wrong grant. A rights reference pointing to a document whose permitted use does not cover the buyer's use passes a mechanical check and fails a legal one.
- Rules that move. Documentation duties such as the EU AI Act's training-content summary for general-purpose model providers and California's AB 2013 can change in scope or detail, which changes the manifest fields buyers ask for.
- Documentation as table stakes. NIST's Generative AI Profile (NIST AI 600-1) names data privacy and intellectual property among twelve risks unique to or exacerbated by generative AI; if buyers come to expect clean manifests from every supplier, manifest discipline protects revenue without earning a premium.
- Concentration. As Reddit's disclosure shows, licensing revenue can rest on two buyers however clean the paperwork.
Manifest questions for any data company, SourceX included
Ask to run the three-way match yourself on deliveries you pick. The same questions apply to SourceX.
SourceX turns records a company chooses to license into AI datasets; along the way it sources and qualifies the supplier, reviews rights, contracts, prepares the data, and handles delivery and payment. SourceX receives no data until the supplying company has accepted the scope and terms in a signed agreement. SourceX does not host large deliveries either; they remain in the supplier's storage or move on encrypted drives. When data does not sit on an intermediary's servers, the manifest is the clearest view an investor gets of what moved.
SourceX has not released delivery, acceptance or revenue figures. Investors receive company and commercial metrics directly, and the current raise accepts only verified accredited investors under Rule 506(c). Questions worth bringing to an investor materials request:
- For three deliveries I choose, can you show the signed scope, manifest, acceptance record and invoice side by side?
- Is each manifest generated by the delivery pipeline at packaging time, and does it carry file checksums?
- Does each manifest line carry a rights reference, and what document does it resolve to?
- What unit does each record count measure, and what are the unique counts after deduplication?
- How many major versions did deliveries go through before acceptance, and where is rework recorded?
Methods note: sources and limits of the three-way match
The three-way match, the five-check table and the manifest evidence request list are editorial frameworks built for this article on 2026-10-27. Their evidence comes from the Croissant 1.0 specification and its responsible AI extension (read in the MLCommons repository), Datasheets for Datasets, Regulation (EU) 2024/1689, California AB 2013, NIST AI 600-1, the Data Provenance Initiative audit and Reddit's quarterly report for the period ended March 31, 2026. Example counts are illustrative.
Not measured: how many data companies keep line-level manifests, how often scope, manifest and invoice disagree, the rework rate at any company, and any SourceX delivery or acceptance figure. The accounts payable comparison is an aid to reasoning, not an accounting standard for data revenue.
Frequently asked questions
What should a dataset manifest include for investor diligence?
Each line should name the record family and source system, the date range, the record count with its unit, unique counts, a rights reference, exclusions with counts, the pipeline version and file checksums. Those fields let an investor match deliveries to signed scope and to billing. Without units and rights references, a manifest shows movement but not meaning or permission.
How is a dataset manifest different from a datasheet?
A datasheet documents a dataset as a whole: why it was made, what it contains, how it was collected and what it suits, as proposed in Datasheets for Datasets. A manifest records one delivery: which files and records went to which recipient, in which version, with which checksums. Investors need both, and the manifest's quality fields should agree with the datasheet.
Can a manifest prove a data company's revenue is real?
Not on its own. It proves what was delivered; revenue quality also needs acceptance against a buyer's written test and invoices tied to accepted lines. The three-way match connects those records, and a mismatch nobody can explain is the finding.
Why do file checksums matter in data company diligence?
A checksum lets anyone confirm that a file is the one the manifest declared. If the hashes on the buyer's receipt match the manifest for the accepted version, delivery is evidenced without trusting either party's account. Croissant strongly recommends recording checksums for all files in a versioned dataset.
Does SourceX share its manifests or delivery metrics?
Not publicly. SourceX has not released delivery, acceptance or commercial figures, and it discusses company metrics with investors directly after a request through sourcex.si/investors. Its offering is limited to verified accredited investors.
Sources
- The Croissant 1.0 specification lets each file object declare a sha256 checksum, lets a record set define a key of fields unique across its records, and recommends semantic versioning in which a major version means existing data was edited, removed or shuffled. Source
- The Croissant responsible AI extension defines properties including rai:dataCollectionMissingData (a description of missing data) and rai:dataBiases (a description of biases in the dataset). Source
- Datasheets for Datasets proposes that every dataset be accompanied by a datasheet documenting its motivation, composition, collection process and recommended uses. Source
- Article 53 of Regulation (EU) 2024/1689 requires providers of general-purpose AI models to publish a sufficiently detailed summary of the content used for training, according to a template provided by the AI Office. Source
- California AB 2013 requires developers of generative AI systems to post documentation on training data, including a high-level summary of the datasets used in development, on or before January 1, 2026. Source
- The Data Provenance Initiative audited more than 1,800 text datasets and reported license omission above 70 percent and error rates above 50 percent on widely used dataset hosting sites. Source
- NIST AI 600-1 (Generative AI Profile) identifies twelve risks unique to or exacerbated by generative AI, including data privacy and intellectual property. Source
- Reddit's quarterly report for the period ended March 31, 2026 states that substantially all of the contract value associated with its licensing revenue is derived from two of its partners. Source
Request SourceX investor materials
SourceX shares company and commercial metrics with investors directly. Any offering is made only to verified accredited investors.