Skip to content

SourceX Research · public sources

Public-source briefings on AI data licensing.

Quick answer

Each briefing summarizes published public sources and lists them at the end; none contains SourceX survey or deal data. The one SourceX research item is the open De-identification Benchmark: its method and synthetic test data are published, and it includes a sample report.

AI Data Price Index

Publicly reported prices, linked to their sources. Not SourceX offers or payouts.

Explore the price index

Benchmarks

Explore SourceX benchmarks
Featured · Open benchmark

The SourceX De-identification Benchmark

How well we protect personal information while keeping business data useful. Includes a sample report.

1,200 synthetic records · 12 entity types · open scorer

Research by Noah Loul.20 briefings

SourceX's own research

SourceX De-identification Benchmark (SX-TXT-1.0) publishes an open method, synthetic corpus and scorer. It includes a sample report. SourceX’s survey and transaction research programs remain in development; no survey or deal findings have been published.

Which questions would research into the buyer value of licensed business records have to answer?

Five questions, each asking whether a licensed dataset changes what a model can do on a business task, and none answered yet. As the note above says, SourceX has published no findings of its own; this section reports no result. It sets out the questions and how each could be tested, so that any later study can be held to them.

Measuring a dataset by holding everything else fixed is established practice. DataComp-LM (DCLM) has participants filter a shared pool or bring their own data, train "with standardized training code and scale-specific hyperparameters", and score the model on 53 downstream tasks "to judge dataset quality". DataPerf, a community-led benchmark suite for datasets, includes a data-acquisition challenge whose buyer must decide what to purchase while it "can observe no sellers' datasets but some summary information from the sellers", which is roughly where an AI team stands before it licenses records.

The second question borrows that framing. How it works lists ten fit signals SourceX evaluates, from volume and historical depth to whether records show how work ended, and labels the illustration "Not a valuation". Whether ratings on those signals track the gain a dataset produces once it is used is a question a study can answer.

Which questions would research into the buyer value of licensed business records have to answer?
QuestionComparisonMeasureResult that would count against value
Does the dataset improve the task it was licensed for?One base model, recipe and compute budget, trained with the dataset or with a matched volume of public or synthetic dataTask score on a test built from later recordsNo gain beyond the spread between repeated runs
Do fit signals rated before delivery predict that gain?Ratings recorded before delivery against the gain measured afterwardsRank correlation with an intervalRatings that fail to order datasets by measured gain
Which parts of a record carry the gain?Full records against copies without outcome fields, linked steps or human feedbackScore change for each removed field groupNo loss when outcomes or links are removed
Does the gain reach companies the model never saw?Tests drawn from suppliers held out of trainingHeld-out score against in-sample scoreGains that disappear on held-out suppliers
How many records does a gain need?Training on fixed fractions of the datasetScore at each fractionA flat curve, where more records add nothing

How would a dataset-utility study be set up so an outsider could check it?

With the plan fixed before any model trains and the report written against that plan. SourceX's methodology page already sets a floor for its own research publications: methodology, sample size and research dates, definitions, limitations and sources, under standards that include "Limitations and conflicts of interest are disclosed" and "Corrections are dated and noted".

Rights draw the outer boundary. For closed deals, section 4 of SourceX's terms describes a license to use the de-identified dataset to "train, validate, test, benchmark, monitor and improve" models operated by SourceX and its customers, and the methodology page says survey and transaction data are aggregated and anonymized, so supplier and buyer identities would not appear in any report. Within that boundary, a utility study would follow six rules:

  • Register first: question, datasets, baseline, test sets, metric and the smallest gain that would count, all dated before training starts.
  • Freeze the recipe: one base model, training recipe and compute budget per comparison, so the data is the only thing that differs.
  • Match the baseline: compare against the same volume of public or synthetic data, never against adding nothing.
  • Split by time and by supplier: test on records dated after the training snapshot and on suppliers left out of training, so leakage cannot pass for value.
  • Repeat runs: several seeds per arm with the spread reported, so a single lucky run never becomes a finding.
  • Version the scoring: DCLM's maintainers note that bug fixes to baseline calculations in September 2025 mean earlier CORE and EXTENDED scores are not directly comparable to new results, so every score should carry the scoring version it was computed under.

What can a reader conclude while no buyer-value results exist?

Nothing about value yet. The questions above have no published answers, and an agenda is a plan, not evidence; it should not be read as a claim that SourceX's datasets improve any model.

Results, when any exist, will carry limits of their own. A gain describes the model, recipe and test it was measured on and may not carry over to another setup. Datasets a buyer agrees to have measured may be the strongest ones. A company measuring the value of what it licenses has a stake in the answer, which is why disclosed conflicts matter. Buyers' tasks can reveal what they are building, so some detail may be withheld, and every withheld detail narrows what a result can prove. A benchmark gain is not revenue, demand or evidence that any dataset will be licensed.

An investor with a specific question about these designs can raise it through the investor page; SourceX shares company and commercial metrics with investors directly, and none is published here.

Methods note. The questions, the comparison table and the design rules were written for this page on 2026-10-27 from SourceX's live research, methodology and How it works pages, its terms, and two public benchmark projects read from their repositories (DCLM at commit 361714b, DataPerf at commit 25f1477). No experiment was run for this page and none is reported; the table's last column states what would count against value, not an expectation of what a study would find.

This article is for information only and is not an offer to sell or a solicitation of an offer to buy securities. Any offering by SourceX is made only to verified accredited investors through its offering documents.

Sources for this section

  • DataComp-LM (DCLM) has participants filter a data pool or bring their own data, train with standardized training code and scale-specific hyperparameters, and evaluate on 53 downstream tasks to judge dataset quality; its maintainers note that baseline fixes in September 2025 make earlier CORE and EXTENDED scores not directly comparable to new ones. Source
  • DataPerf, a community-led benchmark suite for datasets, includes a data-acquisition benchmark in which a buyer decides which datasets to purchase while seeing only summary information from sellers, not their datasets. Source
See if your company qualifies