# Contrarian data thesis: testable questions | SourceX

Canonical: https://refer.sourcex.si/guides/how-to-turn-a-contrarian-data-thesis-into-testable-questions
Last updated: 2026-11-05

Put the consensus view and your contrarian view side by side, write down what each predicts a CEO in your network will say about records the company already keeps, and keep only the questions where the two predictions differ. Ask those about the past, at metadata level, with the answer that counts against you written before the first call. A question that both views answer the same way cannot test anything, however good the conversation feels.

Most of the work is separating what a network can test from what it cannot. Whether AI developers will value a kind of record is a market claim no founder interview settles. Whether the companies you can reach keep that kind of record, own it and would discuss it is a network claim, and that is the part a platform team can actually break.

## Consensus, contrarian and the slice a network can test

A contrarian thesis needs a named consensus to disagree with; otherwise it is an opinion with good posture. In AI data, two layers of consensus are easy to state.

**The market layer.** Epoch AI's 2024 paper projects that, if current trends continue, language models will fully use the stock of public human-generated text between 2026 and 2032 (our paraphrase of its summary). Its later report on AI in 2030 is more relaxed, arguing that scaling can continue for at least a few more years on existing public text and other kinds of data (also a paraphrase). Taken together, the consensus is that public data is finite but workarounds carry the field for now. On the other side, Shumailov and colleagues reported in Nature in 2024 that training indiscriminately on model-generated output causes lasting defects in which the rare parts of the original distribution disappear, and argued that access to original human-generated data matters for learning to continue (our paraphrase). None of these sources mentions business records, buyers or prices.

**The network layer.** Inside venture, a common working assumption is that data advantage lives in data-native startups, and that older operating companies keep records too messy to matter.

An example contrarian thesis, written for this page and not a SourceX claim: the scarce input is records of real work with recorded outcomes, and the companies keeping most of them are mid-sized operators doing that work for themselves, not the startups a fund knows best.

| Part of the thesis | Can CEO conversations test it? | What tests it instead |
|---|---|---|
| AI developers will value records of real work with outcomes | No | Dated primary research, developer documentation and SourceX's own review of fit with current programs |
| Mid-sized operators we can reach keep such records for years in one place | Yes | The question set below |
| Those records are the operators' own, not their clients' | Partly: the CEO's account, pending a rights review | The question set, then SourceX's rights review |
| Their CEOs would discuss licensing selected records | Yes | An ask through the relationship owner |
| Operators keep more of it than the startups in the portfolio | Yes, by comparison | The same questions put to both groups |

Splitting the thesis this way is also what keeps it from turning into a recommendation of any company; [the difference between a category bet and a company endorsement](/guides/the-difference-between-a-category-bet-and-a-company-endorsement) covers the wording side.

## The discrimination test for every candidate question

For each candidate question, fill two cells before you ask it: what the consensus predicts most companies will answer, and what your contrarian view predicts. Keep the question only if the cells differ. An answer counts as evidence for one view only when it is more likely under that view than under the other. If both views expect "yes", a hundred yeses move nothing.

Two more filters follow. The question must be answerable without opening a file: what the records are, roughly how many, how far back, which system, who administers it. And it must ask what the company has done, not what its CEO thinks about AI. Opinions about the future cost nothing to give and tend to flatter whoever is asking.

Three questions that fail the filters turn up in almost every thesis deck:

- "Do you think AI will change your industry?" Both views predict yes, and it asks for an opinion.
- "Would extra revenue from your data interest you?" It leads, and nearly everyone says yes.
- "Is your data valuable?" Nobody can answer it honestly before a review, and the honest answer is "unknown".

This is a different job from [turning approved customer questions into referral content](/guides/how-to-turn-approved-customer-questions-into-useful-referral-content), where the questions come from clients. Here they come from your thesis, and each one exists to risk it.

## The falsifiable interview question set

Eight questions, put to a CEO through the person who owns that relationship, or to the records owner the CEO names. The last column is written before the first call; answers that match it count against the contrarian thesis.

| # | Question | Consensus predicts | Contrarian predicts | Counts against the contrarian view |
|---|---|---|---|---|
| 1 | Since what year has the company kept its own records of each job or case, start to finish, in one system? | Short or scattered histories at operators | Five years or more at operators, longer than at startups | Operators' histories no longer than startups', or split across several tools |
| 2 | Roughly how many such cases does the company close in a typical month? | Startups close more | Operators close more | Operators' monthly volume at or below startups' |
| 3 | Does each record show how the case ended: resolved, paid, reversed or closed with a reason? | Outcomes rarely recorded at operators | Outcomes routinely recorded | Outcomes missing, or kept off-system, at most operators |
| 4 | Is the work done on the company's own account, or on files its clients own? | No difference between the groups | Operators work on their own account | Most operators' records are held for clients |
| 5 | Who administers that system and could describe its fields without exporting anything? | Often nobody inside the company | A named internal owner | No internal owner at most operators |
| 6 | Has anyone approached the company about licensing these records before? | Valuable data has already been approached | Operators have rarely been asked | Most operators have already been approached |
| 7 | Would the CEO take a short conversation about licensing selected records? | Operators decline | At least some operators agree | Every operator CEO declines outright |
| 8 | What was the company's peak full-time headcount, contractors excluded? | The same under both views | The same under both views | Not a test: a gate of 50+ full-time employees at peak, applied before any answer counts |

Question 8 shows the rule at work. Both views agree that a company needs 50+ full-time employees at peak to be introduced, so its answer screens companies in or out and never counts for either side. Question 6 carries a second meaning: under the program terms, a company with any earlier lead, account or email exchange with SourceX is reviewed as a duplicate, so a "yes, SourceX called us" also tells you an introduction from your network would not be credited.

## The evidence each answer stands on

An answer counts only if it came from someone entitled to give it. For every company, record the relationship owner who asked, the CEO who answered or named a records owner to answer, the legal entity that holds the records today and the date. The CEO authorizes any licensing discussion and signs for the company; a functional leader can describe a system but cannot open that conversation.

No answer counts until the company clears the published fit facts: a peak of 50+ full-time employees, with contractors left out of the count, an operating history it can document, the rights to license what it keeps, and someone authorized to speak for it. 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. Nobody asks for samples, exports or screenshots. Three things stay unknown until SourceX reviews a company: whether its records match a current program, whether customer contracts limit reuse, and what any license would cover.

## Sixteen conversations with the losing result written first

The thesis is comparative, so the test needs two groups. Put the same eight questions to eight operators and eight portfolio startups, all past the size gate, all reached through someone who owns the CEO relationship. Before the first call, write the margin that counts as support. A simple rule: the contrarian view gains ground on a question only if at least three more operators than startups give the answer it predicts.

Illustrative. A fictional read-out of sixteen conversations; no real companies.

| Question | Operators matching the contrarian prediction (of 8) | Startups matching it (of 8) | Reading against the margin of 3 |
|---|---|---|---|
| 1. Years of records in one system | 6 | 2 | Supports |
| 3. Outcome recorded on each case | 5 | 4 | No difference |
| 4. Own account, not clients' files | 3 | 5 | Counts against |
| 6. Never approached about licensing | 7 | 3 | Supports |
| 7. CEO would talk | 4 | 4 | No difference |

The read-out neither proves nor kills the thesis; it narrows it. Operators do keep longer histories and are rarely approached, but they record outcomes no better than startups, and more of their records belong to clients. The rewritten thesis is smaller and sharper: operators doing work on their own account. The next round asks only those, which is the point at which [a ten-introduction experiment](/guides/the-ten-introduction-experiment-for-an-ai-data-network-thesis) becomes worth running. For coding answers call by call with stop thresholds, see [how to test a data supply thesis through founder conversations](/guides/how-to-test-a-data-supply-thesis-through-founder-conversations).

## Where a contrarian test fools its author

- **Selection.** Contrarians pick the operators they already admire. Let the relationship owners list candidates before you see the names.
- **Small numbers.** Sixteen conversations show direction, not rates. Report counts, never percentages.
- **The market layer stays untested.** If public text, other kinds of data and synthetic workarounds keep carrying the field, as Epoch AI's 2030 report argues they may for some years, a network rich in operator records could still meet little demand.
- **Rights.** A CEO's account of ownership is not a rights review. Customer contracts and privacy limits are read in SourceX's review and by the company's counsel.
- **Programs move.** SourceX checks relevance to its current programs in review, and a company that misses one program may fit another later. A thesis tested in October can be stale by spring.

Before you treat any result as a verdict on the whole network, run it past [what would disprove your network's data opportunity](/guides/what-would-disprove-your-network-s-data-opportunity), which sorts findings by whether they sink one company, the network or the category.

**Methods note.** The question set and the margin were built on 2026-10-26 from the program terms (version 2026-10-22.2, since replaced by version 2026-11-04), SourceX's published screening criteria and the research summaries cited above. The read-out is fictional. No founder interviews or dated mandate feedback have been published for this thesis, so real figures are not yet available; the method above is how a platform team would produce them.

## 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 an operator's CEO has answered the questions, passed the size gate and asked for the conversation, [register for a partner link](/register), check the facts against [who qualifies](/who-qualifies) and read the [program terms](/terms) before the relationship owner introduces anyone. The [venture platform overview](/for/venture-platform-teams) and the other [partner guides](/guides) cover the steps around it.

Testing a thesis about records is referral work; whether SourceX belongs in a portfolio is a separate investment question, and materials for it are available on request through the [SourceX investor overview](https://sourcex.si/investors).

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

- https://refer.sourcex.si/terms
- https://refer.sourcex.si/who-qualifies
- https://epoch.ai/publications/will-we-run-out-of-data-limits-of-llm-scaling-based-on-human-generated-data
- https://epoch.ai/files/AI_2030.pdf
- https://www.nature.com/articles/s41586-024-07566-y
- https://sourcex.si/investors
