The study needs SourceX's help, because only SourceX knows when a partner's account is approved and whether onboarding has started. It involves no company and no company record, and it changes nothing about how any introduction is reviewed or credited.
Why a baseline of untrained explanations is worth having
For many CEOs, a partner's explanation is the first description of SourceX they hear. If that description promises fit, calls a license a sale or offers privacy assurances SourceX does not make, the CEO starts from a wrong picture that someone later has to correct. Training can only aim at those errors if someone knows which ones come naturally.
A baseline answers that and nothing more. It does not measure whether onboarding works: that takes a comparison over time, like the nine scenario cards scored on day 0 and day 30 in the onboarding experiment, or a chapter-level pilot like the enablement test before a network rollout. How executives read partner material is studied separately in the executive comprehension interview.
Four research questions, written so that no answer is assumed:
- Which elements of SourceX's published description do untrained partners include without being prompted?
- Which misstatements appear, and in how many replies out of all coded replies?
- Do partners who rate themselves confident make fewer misstatements? (Described, never treated as cause.)
- Do elements and misstatements differ by the kind of practice a partner comes from?
| Term | Meaning in this study |
|---|---|
| Explanation | The partner's written reply to the prompt, up to 200 words |
| Element | A point SourceX's own pages state, present in the reply in any wording |
| Misstatement | A claim that contradicts SourceX's pages or the program's rules |
| Absent | An element the reply leaves out; recorded, never scored as an error |
| Before training | After account approval, before any onboarding session, partner material or welcome sequence |
Who writes, and how consent works
The frame is every partner whose account is approved during one entry window of two quarters. Each receives one invitation within two days of approval, from a researcher who works outside partner management, and one reminder.
- Consent covers the purpose, the coding, quotation only with separate permission, and withdrawal up to a dated analysis freeze.
- No consequences. The invitation says plainly that taking part, declining or writing something inaccurate has no bearing on the account, attribution or any reward.
- Excluded before invitation: partners who already began onboarding, returning partners who held an earlier account, and SourceX staff.
- Size. Everyone who consents in the window. If fewer than 60 replies are coded, the report gives code counts only, with no breakdown by group.
- Nothing about companies. The prompt names no company, and coders delete any company name a partner volunteers before coding.
The writing prompt and two follow-up items
Partners see the prompt once, with a 15-minute limit:
A CEO you know well asks you: "What would licensing our data with SourceX actually involve for us?" Write the reply you would send, in your own words, in no more than 200 words. Please do not look anything up while you write.
The prompt mentions no reward, no privacy question and no exclusivity, so that what partners raise on their own can be seen. Two items follow the reply, never before it: how confident the partner is that the reply is accurate (1 to 5), and which of these they had read before writing: SourceX's public site, the program terms, a colleague's description, or none.
The explanation codebook
Each code is defined by meaning, not wording, and each element is keyed to a SourceX page as it read on the day version 1.0 was frozen.
| Code | What it records | Defined as | Keyed to |
|---|---|---|---|
| E1 | What is licensed | Business records the company created itself | SourceX's screening page |
| E2 | The use | AI training and evaluation | SourceX's FAQ |
| E3 | License, not sale | The company keeps its original records | FAQ: "deals are licenses, not sales of your originals" |
| E4 | Exclusive for AI training | The selected data is not licensed to anyone else for that purpose | SourceX's FAQ |
| E5 | No obligation to look | Applying and checking fit commit the company to nothing, and it can decline | FAQ: "Applying and the fit check create no obligation." |
| E6 | Who decides | The company's CEO | The program's rule, on the screening page |
| E7 | SourceX decides fit | Fit is settled in SourceX's own review | SourceX's screening page |
| E8 | The partner's interest | The partner may be paid by SourceX, separately from the company | Section 4 of the terms |
| M1 | Promised fit | Says the company qualifies or will be accepted | Screening guidance is not a promise |
| M2 | Sale framing | Says the company sells its data or gives up ownership | E3's source |
| M3 | Money promised | States a payout, a range or a certainty of payment for the company | FAQ: a profile does not mean an offer or payment |
| M4 | Privacy overpromise | Promises anonymity or no risk of re-identification | FAQ: "No method is perfect" |
| M5 | Named buyer | Names a particular AI lab as SourceX's customer | FAQ: customers described only as AI labs, model developers and research teams |
| M6 | Records requested | Asks the CEO to send files, samples, exports or system lists to the partner | Process page: partners need not export, upload or describe confidential records |
| M7 | Wrong decider | Says a fund, a board, a manager or the partner approves | E6's source |
| M8 | Effortless-money framing | Describes licensing as money that asks nothing of the company | FAQ: preparing a delivery may need someone who can export files |
A reply that coders cannot read as on-topic, or that is blank, gets a disposition in the ledger below, not a code. Keeping it out of the code categories matters for the reliability figures, as the next section explains.
Two coders, blind, with agreement computed once
Coding open-ended answers is common and often reported too thinly to check. A 2024 review by Conry-Murray and colleagues found human-coded open-ended data in 21% of articles across four top psychology journals, with only about a third of those reporting enough to replicate or evaluate the coding. The same authors set out phases for doing it properly (defining the data and units, writing a coding manual and training coders, then establishing reliability), warned that Cohen's kappa can be inflated by testing reliability repeatedly or by changing categories, for instance by adding a missing-data category, and published a preregistration template (our paraphrase of the article's abstract). This design follows them:
- Unit. The whole reply. Each code is marked present or absent for that reply.
- Training on invented replies. Coders practice on 20 replies written by the research team, never on real partners' words.
- Freeze first. Codebook version 1.0, the prompt and this analysis plan are dated and preregistered before any real reply is read.
- Blind and independent. Both coders code every reply without seeing the partner's practice type, confidence rating, reading history or each other's work.
- Agreement once. Agreement is computed per code, a single time, on the full set, and reported whatever it is. It is never recomputed after a code is redefined.
- A third reader. Disagreements go to a third reader, whose decisions are logged with a reason.
- Late codes. A code added after the freeze carries a dated memo and is reported as exploratory.
Counting every partner, including those who wrote nothing
The ledger appears in full in the report, so readers can see who is missing before they read any frequency.
| Ledger line | Count |
|---|---|
| Accounts approved in the window | Not yet counted |
| Excluded before invitation, with reasons | Not yet counted |
| Invited | Not yet counted |
| Declined | Not yet counted |
| No reply after the reminder | Not yet counted |
| Consented, then blank or off-topic | Not yet counted |
| Coded replies | Not yet counted |
| Withdrawn before the freeze | Not yet counted |
Code frequencies are reported as counts over coded replies, in the form "n of N coded replies contained M1", always beside the ledger and never as a share of all partners.
Reporting rules, and the claims the report will refuse to make
As of 2026-10-27, no reply has been collected. When the window closes, the report will give:
- for each code, the number of replies containing it and the agreement figure for that code;
- breakdowns by practice type only where a group has at least 15 coded replies;
- the confidence ratings set beside misstatement counts, described as an association and nothing more;
- quotations only with the writer's separate permission, stripped of anything that identifies them;
- released alongside: the prompt, the consent text, the codebook with its version history, the invented training replies, the ledger and the analysis script.
It will never say that a practice type causes a misstatement, or that any training fixes one. A difference between groups of volunteers is a difference between those volunteers. Partners receive aggregate results only; SourceX's partner team never sees an individual reply.
| Code family | Replies with at least one code | Agreement between coders |
|---|---|---|
| Elements E1 to E8 | Not yet collected | Not yet computed |
| Misstatements M1 to M8 | Not yet collected | Not yet computed |
Ways the explanation study could misread partners
- Writing is not talking. A written reply may be more careful than what a partner says over coffee.
- The prompt shapes the answer. A question about "what it would involve" invites process talk and may under-sample reward or privacy claims.
- Who volunteers. Confident partners may consent more often; the ledger shows the gap but cannot close it.
- The key is SourceX's own wording. A partner who gets a point right in plain words must be credited, which is why codes are defined by meaning.
- Pages change. If a SourceX page changes during the window, replies are still coded against version 1.0's dated key, and the change is reported.
- Prior exposure. Partners who read the public site before writing are not untrained in every sense; the reading item records it.
Checking your own explanation in the meantime
Write your own explanation for a CEO you know, then check it against the eight elements and the eight misstatements above. If it promises fit, calls the deal a sale or offers files to anyone, rewrite it. The one-sentence explanation guide and the two-minute founder outline are tested starting points; other designs sit in the guides hub.
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.
Partners who want their explanation to hold up should register, sign the Channel Partner Agreement before introducing anyone, check each company against the screening criteria and read the program terms in full.