Three kinds of statement appear below, and they are labeled. Known: published research on data supply and evaluation, and SourceX's published fit criteria. Conditional: whether any company you introduce is accepted, which only SourceX's review decides. Proposed: the thesis itself, treated here as a hypothesis to test.
Turn the category thesis into a company question
Begin with a thesis a partner could actually lose. A working example, not a finding:
Thesis (proposed). AI developers will keep needing real records of multi-step business work in areas where no automated check can grade the output: claims decisions, dispatch exceptions, underwriting referrals.
The evidence cuts both ways, which is why it is worth testing.
- For. Epoch AI estimates the effective stock of quality-adjusted public human text at around 300 trillion tokens, and gives an 80% confidence interval that it will be fully used at some point between 2026 and 2032 if current trends continue. The GDPval benchmark builds each task from actual work product created by an expert professional across 44 occupations, and grades mainly by expert comparison because automatic grading is hard.
- Against. Epoch AI's outlook to 2030 argues data is unlikely to be the binding limit on scaling, with synthetic and multimodal data closing the gap. The International AI Safety Report 2026 notes that in domains with verifiable outputs, such as mathematics, programming and formal reasoning, models can generate their own training data by proposing solutions and checking them.
Read together, the evidence leaves the thesis standing only in a narrow form: judgment that cannot be auto-checked. That narrowing is the useful part, because it converts a market view into a trait you can ask a CEO about:
Company question. Which companies we can reach hold at least five years of records in which a person made a judgment call no automated test can grade, captured together with the outcome, in a system the company controls?
Broader market sizing belongs on sourcex.si; for how AI data spend reads from a private equity seat, see what AI data company growth signals.
The thesis-to-network activation canvas
Six boxes, filled in order, on one page. The example column is Illustrative.
| Box | What it answers | Example (Illustrative) | Evidence standard |
|---|---|---|---|
| 1. Thesis | One sentence that could be false | Developers keep needing real records of hard-to-grade business judgment | Names a mechanism and a domain |
| 2. What would disprove it | The observation that ends it | Synthetic data succeeds in hard-to-grade domains; thesis picks fare no better than controls | A cited primary source or a pilot measure |
| 3. Company question | An observable trait | Five or more years of judgment records with outcomes, in one company-controlled system | Answerable from metadata alone |
| 4. Ownership and relationship evidence | Who owns the records, who authorizes, who owns the CEO relationship | Company-owned records; CEO reachable through a named partner | Confirmed by the relationship owner, never by samples |
| 5. Narrow pilot | The smallest test that could fail | 8 thesis-selected companies, 4 controls, six weeks, consented introductions only | Fixed before the first call |
| 6. Keep or drop rule | The result that changes the thesis | Drop if thesis picks pass the fit facts and gain CEO consent no more often than controls | Written down in advance |
Method note. The canvas combines SourceX's fit criteria, taken from the program terms and screening page as published on 2026-10-26, with the research cited above. It measures fit facts, CEO consent and team hours. It does not measure buyer demand, record quality or payout, none of which can be known before SourceX's review, and it infers no buyer relationships.
| Term | Meaning on this page |
|---|---|
| Thesis | A one-sentence claim about where AI developers will need real records, written so it could be false |
| Company question | The thesis restated as a trait a CEO can confirm from metadata |
| Control group | Companies picked without the thesis, to test whether the thesis improves selection |
| Kill rule | The result, written before the pilot, that ends or rewrites the thesis |
What ownership and relationship evidence does each company need?
Box 4 is where most theses quietly fail, because a trait can belong to the wrong owner.
| Evidence | Why it matters | Where it comes from |
|---|---|---|
| The company, not its clients, owns the records | Client-owned material rarely passes SourceX's screen | The CEO or system owner, in a metadata conversation |
| The CEO can authorize the discussion | The CEO authorizes and signs for the company | Your relationship owner |
| 50+ full-time employees at peak, contractors excluded | A published fit fact; placed staffing workers do not count | Board materials, the CEO or public filings |
| Documented operating history | A published fit fact | System name and years, never exports |
| Group structure | One reward cap covers a parent and every subsidiary | Ownership chart |
Location is the last fit fact. Companies in the U.S., Canada and the EU qualify, as do companies in the UK, Australia, New Zealand, India, Singapore, South Africa, Latin America, the UAE and Saudi Arabia when at least 70% of the data is in English. Two answers arrive only with SourceX's review: whether the records suit a program SourceX is running now (criteria differ by program, so today's no can become a later yes), and whether customer contracts restrict reuse. No introduction can be submitted before the partner signs the SourceX Channel Partner Agreement, earlier named the Channel Partner Non-Circumvention and Non-Disclosure Agreement.
Run a small test, and say first what would change the thesis
Design. Choose 12 reachable companies: 8 selected by the company question and 4 selected the way your team normally would. For each, the relationship owner asks the CEO a short metadata question set (which system, how many years, roughly how many records, who owns them) and then whether they want a conversation. Introduce only with consent, through your partner link, and log hours. Nobody on the team sees a file, sample or login at any point.
Measures. All counts or durations: companies passing the four fit facts on metadata; CEOs who agreed to talk; hours per company; and any review status a company chooses to share back.
Decision, set in advance.
- Keep if the thesis group beats the controls on both fit facts and CEO consent by the margin you wrote down.
- Rewrite if it wins on fit facts but not consent: the thesis finds the records but not willing owners.
- Drop if there is no difference.
How to read a result (Illustrative counts that show the reading, not a result):
| Group | Companies | Passed four fit facts | CEO agreed to talk | Reading |
|---|---|---|---|---|
| Thesis picks | 8 | 5 | 3 | Selection works; consent is the bottleneck |
| Controls | 4 | 1 | 1 | Usual judgment finds fewer fits |
Twelve companies prove nothing statistically. The point is to retire a weak thesis cheaply before it consumes founder attention. SourceX has not published results from any such pilot, so treat the table as a reading guide only.
A counterexample: the trait is real, the owner is wrong
Illustrative. A 300-person vertical software company in the portfolio runs claims workflows for insurers, and its database holds years of adjuster decisions with outcomes. It answers the company question perfectly and fails box 4: the records belong to its customers. The fix is not to push the founder. It is to rewrite the company question toward the insurers and claims administrators themselves, whose CEOs the founder knows, which moves the work outside the portfolio. The five-ring network map shows how to find who owns those relationships.
Risks that could make the thesis wrong
- Synthetic data. Epoch AI's 2030 outlook and the International AI Safety Report both describe paths where generated data substitutes for human records, at least in checkable domains. If that spreads to judgment work, the thesis weakens.
- Rights. Hard-to-grade records can sit in client-owned systems, as the counterexample shows.
- Demand by program. SourceX applies each program's criteria, so a well-chosen company can still be declined today.
- Governance shifts. NIST released AI RMF 1.0 on January 26, 2023 as a voluntary framework and says it is being revised; changes in how developers are expected to govern AI could change which records they want and how much documentation travels with them.
- Endorsement drift. A category bet is not a company endorsement. Keep founder conversations about records, not about whether the company is a good investment.
The next six weeks. Fill the six boxes this week, pick the 12 companies, and run for six weeks. Record the keep, rewrite or drop decision with the counts that produced it. Program basics for funds sit on the venture platform teams page, and how platform teams explain data partnerships helps with the founder conversation.
Next step
Refer a company. No cap on rewards. Default reward on a company's first qualifying transaction: the greater of $50,000 plus 5% and 15% of Actual Supplier Payout, with no cap; further qualifying transactions first agreed within 12 months of the first Closing pay 15%. Eligibility and payment conditions apply, including the USD 100,000 Minimum Transaction, and a signed Schedule A or addendum can differ. A lead, meeting or signature alone does not trigger payment.
When a thesis pick has a confirmed relationship owner and a CEO keen to talk, set up your partner link, measure the company against the who-qualifies page and work through SourceX's referral terms before the introduction goes out.
Readers weighing SourceX itself as an investment should treat that as a separate question from this pilot and can request materials through the SourceX investor overview.
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.