Skip to content

AI data market

A dataset acceptance infrastructure business sells pass marks, not checks

By SourceX Editorial · Reviewed by Noah Loul ·

10 sections

Short answer

A dataset acceptance infrastructure business becomes a product only where one published pass mark serves many buyers, an independent party runs the test, and the test repeats. The checking code is a commodity: 47 of 60 core Great Expectations checks wait for a user-supplied value such as a bound, set or pattern, and the engines are free or built into clouds. Buyer-specific rules are services; shared standards with re-testing are the product.

Checklist

Dataset acceptance

This page covers A dataset acceptance infrastructure business sells pass marks, not checks.Illustrative

Key takeaways

  • Acceptance work has three layers: an engine that computes checks, rules that set pass marks, and a standard with an independent tester; the third has the clearest product economics.
  • 47 of 60 exported Great Expectations core checks and 23 of 27 listed Deequ rule types need a value the user supplies, so the buyer-specific part is exactly what open tools leave blank.
  • The engines are Apache 2.0 and AWS runs Deequ as a managed service; Soda Core's Elastic License 2.0 bars hosting it, showing vendors defend the hosted layer.
  • Certification became recurring through surveillance and impartiality rules: UL Solutions books ongoing certification over time, and ISO/IEC 17065 and SEC Rule 17g-5 police independence.
  • Ask any acceptance seller, SourceX included, what share of a new customer's rules were reused and who pays for the verdict.

Can a dataset acceptance infrastructure business sell a product, or only a service?

They can be a product only where one published pass mark serves many buyers, an independent party runs the test, and the test has to be repeated. Without those three conditions, acceptance work is either free software or labor sold one buyer at a time.

An acceptance standard here means the written rules a delivered dataset must pass before the buyer takes it: completeness thresholds, allowed values, schema, label agreement, freshness. A data company may pitch the tooling and the rules together as a platform. The investor question is which part a customer would pay for twice, and which part it could get free or would rather write itself.

This page splits acceptance work into three layers, counts how much of two widely used open-source check libraries depends on values the user must supply, and looks at the industries where testing to a standard did become a recurring business.

Three layers of acceptance work, and who supplies each

Every acceptance decision combines an engine that runs checks, a set of rules that says what passes, and a party whose verdict others rely on. Price the three separately:

  • The engine: code that computes completeness, uniqueness, ranges, distributions and schema matches. It is reusable across buyers by design.
  • The rules: the numbers, sets, patterns and thresholds that turn a computed metric into pass or fail. They usually come from one buyer's task and change when the task changes.
  • The standard and its tester: a rule set published before any test, applied by a party independent of whoever pays, and accepted by more than one relying party. Certification businesses are built on this layer.

47 of 60 checks wait for someone else's number

Most built-in checks are templates. In Great Expectations at commit 99f4d0c (22 October 2026), 60 core expectations are exported, and 47 of them, 78.3 percent, declare a field for an expected value: a minimum or maximum, a value set, a pattern, a type, a schema, a reference table or query, or a threshold. The other 13 test a property of the named columns only, such as being non-null, unique, parseable, ordered or equal to one another.

Deequ, an open-source library published under AWS Labs, has the same shape. Its README lists 27 rule types for its rule language, DQDL, once the entry for combining rules is set aside; 23 of the listed examples carry a threshold or expected value, such as completeness above 0.9 or a row count below a limit, and 4 are yes-or-no checks: complete, unique, primary key and column exists.

The counts show where the decision lives. The library computes the metric, but the value that makes a dataset pass is written by whoever authors the suite. For a buyer accepting licensed data, that value comes from its own task, so the part of acceptance that differs from buyer to buyer is precisely the part the open tools leave blank.

Two unit charts: 47 of 60 exported Great Expectations core checks and 23 of 27 listed Deequ rule types need a user-supplied bound, set, pattern or threshold; the rest check a property only.
Great Expectations commit 99f4d0c and Deequ commit d7f7783; counts from the article's script and stated rules.

The check engine is already free, and clouds absorb it

Both libraries carry the Apache 2.0 license at the commits counted, so anyone can run the engine without paying. AWS goes a step further: its documentation describes AWS Glue Data Quality as a managed, serverless service built on top of the open-source Deequ framework, with rules written in DQDL. A company selling a check engine competes with free code and with a feature of the cloud its customers already rent.

Vendors defend the hosted layer instead. Soda Core's license at commit b3aba21 (27 October 2026) is the Elastic License 2.0, whose limitation reads: "You may not provide the software to third parties as a hosted or managed service, where the service provides users with access to any substantial set of the features or functionality of the software." That clause protects the vendor's own hosted product, and it tells an investor where the vendor expects value to sit: in operating checks for customers, not in the checks themselves.

Single-buyer rules turn acceptance into services

When every buyer writes its own pass marks, the acceptance provider sells hours. Rules for a field-service dispatch dataset share little with rules for a coding benchmark, and each new buyer brings a new suite, a new negotiation over thresholds and new disputes about edge cases. The revenue ends when the project does.

That is the strongest alternative explanation for an acceptance platform's growth: the platform label may be covering a services business. The test is arithmetic. Ask what share of the rules in a new customer's suite already existed in an earlier customer's suite, how many hours a suite takes to write and maintain, and whether those hours scale with buyers or with datasets. A product shows reuse rising with each customer; a service shows hours rising with them.

Where testing to a standard became a recurring business

Product certification is the clearest precedent. UL Solutions' annual report for 2024 describes ongoing certification services that confirm certified products stay compliant, including factory inspections that check production matches the tested configuration, and says revenue from those compliance programs is recognized over time as customers receive the benefit of continued certification. The recurring part is the surveillance, not the first test.

Independence is what makes the verdict worth buying. ISO/IEC 17065 sets requirements for the competence, consistent operation and impartiality of bodies that certify products, processes and services. Credit ratings show a regulator putting a number on it: SEC Rule 17g-5 lists conflicts of interest that rating agencies must manage or avoid, and one agency's filing describes monitoring "the risk of exceeding the 10% threshold of total net revenue in accordance with Rule 17g-5(c)(1)", the paragraph that bars rating for a client that large.

Carried over to data, an acceptance business becomes a product when the rule set is fixed and published before testing, the tester is independent of the supplier under test and of any single buyer, several buyers accept the same certificate, and each refresh or new version must be re-tested to keep it. None of the sources reviewed for this page shows a widely adopted pass-mark standard for AI training data; that gap is both the opening and the risk.

Five ways to sell acceptance, compared per buyer and per standard

Normalize each model to the unit that repeats, so a count of contracts cannot pass for reuse. The rows describe structures and do not rank companies:

Five ways to sell acceptance, compared per buyer and per standard
ModelWho writes the pass marksWhat recursMetric normalized for comparisonMain risk
Open-source check engineWhoever authors the suiteNothing billable; upkeep is donated or vendor-fundedActive users per maintainerFree substitutes and cloud bundling
Hosted data-quality softwareThe customer's own teamSeats or usage of the hosted serviceRevenue per monitored dataset, services split outCompeting with free engines and cloud features
Buyer-specific acceptance testingEach buyer, one project at a timeOnly another project from the same buyerHours per suite and share of rules reused across buyersServices economics under a platform label
Independent certification to a published standardA scheme owner, before any testSurveillance and re-testing to keep the certificateRelying parties per certificate and re-tests per yearIndependence conflicts and slow adoption
Licensor's acceptance gate inside its dealsThe licensor, under each deal's termsEach new delivery or onward licenseShare of checks reused across deals and datasetsCost carried per deal; value visible only in close rates

Applied to SourceX: acceptance sits inside each license

SourceX is a labeled application of the question, described only from its published pages. It does not sell acceptance testing as a separate product; acceptance is a step inside its licensing deals. Its terms say it reviews delivered data and closes only once it has accepted the data in writing and every condition is met; data that does not close earns no payment, and declined data is deleted.

The reusable layers it describes are the contract and the rights work. Its investors page cites one standardized licensing agreement used across deals and says ownership checks and removal of personal details are built into the workflow. That puts SourceX in the last row of the table: the gate's cost lands on each deal, and its value shows in whether deals close and datasets are licensed onward.

How often its acceptance checks are reused across datasets and buyers is the measure that matters here, and SourceX has not published it, nor acceptance or close rates. It shares company and commercial metrics with investors directly, and materials can be requested through its investors page.

What could make acceptance standards a weak business

Five developments would weaken the case, and each can be watched:

  • Free engines: Great Expectations and Deequ are Apache 2.0 at the commits counted, and AWS runs Deequ as a managed service, so the check layer earns little on its own.
  • Private, buyer-specific tasks: if each buyer's pass marks stay unpublished and different, reuse never rises and margins stay at services levels.
  • Independence conflicts: a tester paid by the suppliers it tests faces the same conflict rating agencies are regulated for, and buyers may discount its verdicts.
  • Third-party license terms: Soda Core's Elastic License 2.0 restricts offering it as a hosted service, a reminder that a business built on another firm's engine depends on terms that firm sets.
  • Slow adoption: certification pays only when relying parties accept the certificate, and no adopted pass-mark standard for AI training data appears in the sources reviewed.

Counting rules, sources and diligence questions

Great Expectations: repository great-expectations/great_expectations at commit 99f4d0cca81de00db3af8859e16127e9b9927dae (22 October 2026), parsed with Python's ast module without importing or running it. Population: the 60 classes that the package initializer of great_expectations/expectations/core imports; four files in that folder are not imported (each raises NotImplementedError) and are left out. A class takes an expected value if its own annotated fields include min_value, max_value, value_set, value_pairs_set, value, threshold, type_, type_list, like_pattern, like_pattern_list, regex, regex_list, json_schema, strftime_format, partition_object, quantile_ranges, sum_total, column_set, column_index, other_table_name, base_query, comparison_query, unexpected_rows_query or multiplier, or column_list in a class whose name starts with ExpectTableColumns. Optional fields count; inherited fields do not. Own fields left uncounted as scope or behavior settings: allow_relative_error, bucketize_data, column, column_list elsewhere, comparison_data_source_name, condition_parser, double_sided, escape, exact_match, ignore_row_if, internal_weight_holdout, match_on, method, mostly, or_equal, row_condition, strict_max, strict_min, strictly, tail_weight_holdout and ties_okay.

Deequ: README.md at commit d7f778310cadcfaf7ad641bdfe07368416cc7c64 (16 September 2026), entries under Supported DQDL Rules, with indented sub-bullets joined to their entry and Composite Rules left out; an entry takes a threshold if its text contains <, >, =, >=, <=, between or in [. Licenses come from each repository's LICENSE file. Certification and ratings evidence comes from search excerpts of UL Solutions' 2024 annual report, the eCFR, an Egan-Jones filing and ISO's catalog page, checked 27 October 2026. Not measured: any acceptance vendor's revenue, rule reuse inside any company, or adoption of any data standard. A reviewer can rerun the count with the saved script.

Questions for any company selling acceptance, SourceX included:

  • What share of the rules in the newest customer's suite were reused from earlier customers?
  • Who writes the pass marks, and who pays for the test?
  • Do any buyers accept a certificate or report without re-running the tests themselves?
  • What triggers a re-test: a refresh, a new version or a fixed schedule?
  • How many hours does a new suite take, and do hours grow with buyers or with datasets?

Frequently asked questions

What is dataset acceptance infrastructure?

It is the tooling, rules and process a buyer uses to decide whether a delivered dataset meets the agreed bar before taking it. It has three layers: an engine that computes checks, rules that set the pass marks, and sometimes a standard with an independent tester whose verdict others rely on.

Is data quality tooling a good business for a data company?

The engine alone is hard to charge for, because widely used libraries such as Great Expectations and Deequ are Apache 2.0 and AWS offers Deequ as a managed service. Value tends to sit in hosting and operating checks, or in a published standard that many buyers accept.

Why does independence matter for acceptance testing?

A verdict is only worth buying if the tester has no stake in the result. ISO/IEC 17065 sets impartiality requirements for product certification bodies, and SEC rules bar a rating agency from rating for a client that supplied a large share of its revenue. A data tester paid by the suppliers it grades faces the same conflict.

Can open-source data validation tools replace a buyer's acceptance test?

They can run it, but they cannot write it. In the count on this page, 47 of 60 core Great Expectations checks and 23 of 27 listed Deequ rule types need a value the user supplies, so the pass marks still come from the buyer's task.

How could a data acceptance certification earn recurring revenue?

The way product certification does: through surveillance. If every refresh or new version of a dataset must be re-tested to keep a certificate that several buyers accept, testing repeats by design. Without re-testing or shared acceptance, it is a one-off service.

Sources

  • The core expectations Great Expectations exports at commit 99f4d0c, counted by the article's script: 47 of 60 declare an expected-value field. Source
  • Great Expectations is licensed under the Apache License 2.0. Source
  • Deequ's README lists the supported DQDL rule types with an example of each; 23 of 27 examples carry a threshold or expected value. Source
  • Deequ is licensed under the Apache License 2.0. Source
  • Soda Core is licensed under the Elastic License 2.0, which bars providing the software to third parties as a hosted or managed service. Source
  • AWS describes AWS Glue Data Quality as a managed, serverless experience built on top of the open-source Deequ framework, with rules in DQDL. Source
  • UL Solutions' 2024 annual report describes ongoing certification services, including factory inspections, with compliance-program revenue recognized over time. Source
  • ISO/IEC 17065 contains requirements for the competence, consistent operation and impartiality of product, process and service certification bodies. Source
  • SEC Rule 17g-5 sets out the conflicts of interest that rating agencies must disclose and manage and those that are prohibited. Source
  • An Egan-Jones filing describes controls to monitor the risk of exceeding the 10% threshold of total net revenue under Rule 17g-5(c)(1). Source
  • SourceX's published terms: it reviews delivered data, closes only after accepting it in writing, pays nothing for data that does not close and deletes declined data. Source
  • SourceX's investors page cites a standardized licensing contract and built-in rights and privacy tooling, and says it shares metrics with investors directly. Source

Request SourceX investor materials

SourceX shares company and commercial metrics with investors directly. Any offering is made only to verified accredited investors.

See if your company qualifies