How to test a data supply thesis through founder conversations: three predictions written before the first call, a coding sheet for every answer and checkpoints that can end the test early

Short answer

Split the thesis into three predictions a founder can confirm or contradict: companies in your segment keep years of work records, those records are the company's own, and the CEO wants a licensing conversation once it is explained. Write down, before the first call, how many contradictions end each prediction. Then hold short conversations in blocks of four, ask about systems and ownership before licensing comes up, code each answer against the predictions and read the tally at every checkpoint. Stop the moment a written threshold is crossed.

Visual briefFull-size diagram
A dot tally of three thesis predictions across twelve planned founder calls in blocks of four; after eight calls the ownership prediction reaches its third contradiction at the second checkpoint, so the thesis is rewritten toward the record owners and the last block is never run.
Illustrative tally. Predictions, thresholds and codes follow the protocol on this page; no real companies or results.

A thesis tested only on founders who already like you will survive almost anything. The protocol below gives it a fair chance to fail, and nobody running it ever sees a record, a file or a login.

Split the category view into predictions a founder can contradict

A category thesis is too large for any founder to answer. Here is a working example, not a finding:

Thesis (proposed). As more of the public web closes to AI crawlers, developers will need more licensed operating records, and established services companies with 100 to 500 employees hold more of those records, in systems they control, than young software companies do.

Published research supports the first half and complicates it. The Data Provenance Initiative audited about 14,000 web domains that feed AI training corpora and found that in a single year, 2023 to 2024, robots.txt restrictions made about 5% of all tokens in the C4 corpus, and 28% of its most actively maintained critical sources, off-limits to AI use; terms-of-service restrictions covered about 45% of C4. Against that, Microsoft's Phi-4 technical report says synthetic data makes up most of that model's training data, generated from seeds that include excerpts of web pages, books and code.

Neither paper says whether a 200-person claims administrator owns ten years of adjuster notes or would license them. That company-level half of the thesis is the only part founder conversations can test, so restate it as three nested predictions:

PredictionA founder answer that supports itA founder answer that contradicts itThreshold, fixed before the first call (Illustrative)
P1. Records exist: segment companies keep five or more years of work records in one system"Our job system goes back to 2014""We moved systems in 2022 and kept nothing older"Stop at 4 contradictions in the first 8 calls
P2. Ownership: the records are the company's own, not its clients'"Those are our own dispatch notes""That sits in our clients' accounts under their contracts"Stop at 3 contradictions among segment companies
P3. Willingness: the CEO wants a licensing conversation"Send me the details""No outside use of our records, ever"Stop if fewer than 2 CEOs say yes across 12 calls

Method note. The thresholds are our judgment for a 12-call test, chosen so that a weak thesis dies within two blocks. Change them before the first call if you like; never after. The predictions use SourceX's published fit criteria (terms and screening page, checked 2026-10-26) and the two papers above; they measure what founders say about their own systems, not buyer demand, record quality or payout.

The interview-led thesis test protocol

Seven steps, each with the rule that stops the test from drifting toward the answer you want.

StepWhat happensThe rule that keeps it honest
1. Date the predictionsOne page with P1 to P3 and their thresholds, sent to a colleague who will not run callsNo edits once the first call is booked
2. Pick twelve companiesNine inside the segment and three contrast companies you expect to fail P1 or P2, such as software firms whose product data sits in customer accountsDraw from your whole network list and write down why each company was chosen
3. Ask through the relationship ownerWhoever owns the CEO relationship asks for twenty minutesA decline stays in the count as a decline; it is not replaced with a warmer founder
4. Hold the conversation in a fixed orderOperations, then records, then ownership, then licensingThe thesis is never mentioned; licensing and any reward come up only in the last question
5. Code the same dayFor each prediction: supports, contradicts or no evidence, with the founder's metadata answer quotedWhere possible a second person codes from the notes without knowing which group the company is in
6. Read the tally every four callsCompare counts with the thresholdsContinue, rewrite or stop exactly as step 1 says
7. Log dated mandate feedbackFor each CEO who chose an introduction, record every review status SourceX shares and its dateA status counts as evidence only once the review has ended

The twenty-minute conversation. The questions describe records without touching them.

  1. "Which system does the core work run through, and since what year?" (P1)
  2. "Roughly how many records a year go into it, and was anything lost in a migration?" (P1)
  3. "Whose information is it? Do customer contracts say how you may use it?" (P2)
  4. "If someone asked to license some of it, who would decide?" (authorization)
  5. Only now, two scripted sentences: some AI developers license real business records through companies such as SourceX, which handles rights review, preparation and payment; would you want a conversation about it? If the person asking is a registered partner, they say here that they may receive a referral reward. (P3)

Record the last answer as yes, not now or no, with the reason in the founder's words. The founder explanation guide covers the longer conversation that follows a yes.

A coding sheet after two blocks (Illustrative). Generic companies, invented answers.

CallCompany (Illustrative)GroupP1 recordsP2 ownershipP3 willingness
1240-person HVAC services groupSegmentSupports: dispatch system since 2013Supports: own work ordersYes
2130-person claims administratorSegmentSupports: claim files since 2011Contradicts: files held under carrier contractsNot asked
3310-person freight brokerSegmentSupports: load records since 2015SupportsNo: "not while we refinance"
4Young analytics startupContrastContradicts: founded in 2023Not askedNot asked
5160-person IT services firmSegmentSupportsContradicts: ticket data lives in client tenantsNot asked
6420-person facilities contractorSegmentSupportsSupportsYes
7200-person billing services firmSegmentSupportsContradicts: client billing recordsNot asked
8180-person software companyContrastSupportsContradicts: product data in customer accountsNot asked

At checkpoint 2, P1 has one contradiction, from a contrast company, and holds. P2 has three contradictions among segment companies and crosses its line. The test stops after eight calls, which is the point of a threshold: four more calls would only have spent founders' goodwill on a question already answered.

The relationship and ownership evidence behind each call

Every row needs three kinds of evidence before anyone hears the word licensing.

  • Relationship. A named person who owns the CEO relationship and has agreed to ask. Funding status plays no part; an LP-owned distributor belongs on the list as much as a portfolio company.
  • Authorization. The CEO decides whether the company discusses licensing and signs for it. An introduction the decision maker has not heard about is usually not a fit, which is why question 5 goes to the CEO, not to an operations lead.
  • Ownership. Records the company created itself. Material that belongs to its clients is usually not a fit; the ownership questions guide keeps that check at metadata level.

The published fit facts are the floor for any row that turns into an introduction: a full-time headcount that peaked at 50 or more, contractors excluded, a documented operating history, the right to license the records and an authorized contact. 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. The USD 50,000 floor applies once per referred company, so a parent and its subsidiaries treated as one company share that one floor; there is no cap on rewards.

Ask one more question before any introduction: has the company already been in touch with SourceX? The terms count any earlier lead, account or email correspondence as an existing relationship, so such a company is reviewed as a duplicate and is not credited to a later partner. Three things stay unknown until SourceX reviews a company: whether its records match a current program (criteria differ by program, so a no today can become a yes later), whether customer contracts restrict reuse and what a license would cover. Whoever submits must first have signed the SourceX Channel Partner Agreement, once titled the Channel Partner Non-Circumvention and Non-Disclosure Agreement.

What each result does to the thesis

Write this table into the step 1 page so the decision is made before the evidence arrives.

Result at a checkpointWhat it meansWhat changes
P1 crosses its thresholdThe segment does not keep the history you assumedRedefine the segment by age or workflow, or drop the thesis
P1 holds, P2 crossesThe records exist but belong to the segment's clientsRewrite the thesis toward the record owners themselves
P1 and P2 hold, P3 falls shortRecords and rights exist; CEOs do not want the conversation nowKeep the supply claim, stop asking and revisit on dated events such as a system migration
Contrast companies score like the segmentThe segment definition adds nothingDrop it and screen on the published fit facts alone
All three hold after twelve callsThe thesis survived your network, which is not proofContinue with consented introductions and wait for dated review statuses

The sample run above lands in the second row. The rewritten thesis points at carriers, hospital systems and the other clients whose records the segment companies process, and a new page of predictions starts the next test. For the step before any of this, deciding which ten conversations a fund should hold first, see the first ten conversations for a fund; for a portfolio-wide version with control companies, see the thesis-to-introduction canvas.

Where founder conversations mislead

  • Warmth. Founders who like the asker say yes to conversations they never take. P3 counts a yes only when the CEO agrees to an introduction, not to "keep me posted".
  • Leaking enthusiasm. An asker who believes the thesis hints at it. The fixed order and the scripted licensing sentences limit that; a second coder catches some of the rest.
  • Small numbers. Twelve calls can retire a weak thesis cheaply. They cannot prove a strong one.
  • Reach. You can only call companies someone in your network knows, so a thesis can be right about the market and still fail on your list.
  • Shifting supply. If synthetic data keeps displacing human records, as the Phi-4 report suggests it can for some training, demand for operating records could narrow. SourceX checks relevance to each program in its review, and no founder can answer that question for you.
  • Lagging feedback. Review statuses arrive weeks or months after an introduction. Judge the thesis on the coding sheet first and treat statuses as a later check.
  • No published benchmarks. SourceX has not published results from founder-interview tests or conversion figures for introductions, so a tally can be read only against the thresholds you wrote, never against an industry norm.

This week. Write the three predictions and their thresholds on one page, pick the twelve companies with the reason for each and book the first block of four. Introductions are a separate, later decision that belongs to each CEO; an introduction is never an endorsement, and introducing without endorsing explains the difference. Program basics for funds are on the venture platform teams page.

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 CEO from your coding sheet says yes to the conversation and SourceX has not heard from the company before, sign up as a partner, recheck the company on who qualifies and read through the program terms before the introduction goes out.

Testing a thesis in your own network and evaluating SourceX as a company to invest in are separate questions; for the second, 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.

  1. Step 1Share your linkSend your personal link to a company you know.
  2. Step 2Company appliesThe company applies itself at /apply.
  3. Step 3Company signs and is paidThe company signs, delivers its data and is paid by SourceX under its agreement.
  4. Step 4You get your rewardthe greater of $50,000 plus 5% and 15% of the actual supplier payout, with no cap. A reward becomes payable after SourceX has paid the referred company under its signed agreement and the payout conditions in the program terms are met, including finance approval and your tax and payment details.

Common questions

Should founders be told that a thesis is being tested?

Tell them plainly that you are learning how companies like theirs keep operating history, that you will not ask for files and that a question about licensing comes at the end. Leave out your hypothesis itself, because naming it steers the answers toward it.

Does a founder conversation count as an introduction to SourceX?

No. Credit is reserved for the partner who first validly introduces a company SourceX does not already know, and it counts only when that company submits a verified application inside the 90-day window. A conversation with you is research; an introduction follows only when the CEO wants one and the partner has signed the Channel Partner Agreement.

What if the founder has already been in touch with SourceX?

Record it and do not introduce the company. The terms count any earlier lead, account or email correspondence with SourceX as an existing relationship, so a later introduction is reviewed as a duplicate. The answers still count as evidence for or against the thesis.

Sources

  1. refer.sourcex.si/terms
  2. refer.sourcex.si/who-qualifies
  3. arxiv.org/abs/2407.14933
  4. arxiv.org/abs/2412.08905
  5. sourcex.si/investors

Free resources

Know a company with valuable proprietary data?

Become a referral partner, get your link and introduce the company’s CEO (in an owner-run company, usually the owner). Referral partners can join from most countries. We can't work with partners in countries or regions under U.S. sanctions or on U.S. restricted-party lists. The companies you introduce should be based in the U.S., Canada and the EU, or in the UK, Australia, New Zealand, India, Singapore, South Africa, Latin America, the UAE and Saudi Arabia with at least 70% of their data in English.

Refer a company →

I own a business

Explore licensing your company's data directly to SourceX for AI training and evaluation. Start a short assessment; no uploads needed.

Start an assessment