This essay is for a VC platform strategist whose fund has started telling a category story in LP letters, essays or a partner meeting deck. The program facts are drawn from SourceX's terms (Version 2026-10-22.2, since replaced by version 2026-11-04) and screening page as read on 26 October 2026. The story, the sorting and the pilot results below are Illustrative and use this page's own method.
The story as written, and the claims folded inside it
Category stories travel because they are compressed. One paragraph carries a claim about AI developers, a claim about operating companies, a claim about what nobody has done yet, a claim about the fund's reach and a claim about the future. Each needs a different kind of evidence, and only two of them can be observed through a company introduction.
Private equity already treats its exit narrative this way: every claim in it gets tested by someone, as the exit story and value creation plan guide describes. A venture category story deserves the same treatment before it reaches founders. Here is an Illustrative version, of the kind a platform team might draft:
- AI agents will be judged on real work, not on puzzles.
- The record of real work lives inside established operating companies, not on the open web.
- Most of those companies have never been asked about it.
- Our network reaches their CEOs.
- A new category forms around that supply, and we are early to it.
The narrative-to-experiment plan
The plan gives each sentence a home. Sentence 1 is a claim about how developers evaluate models, and the evidence for it lives in research, standards and developer documentation. NIST's AI Risk Management Framework is a useful example of what such evidence does and does not say. Its Map 2.3 subcategory reads "Scientific integrity and TEVV considerations are identified and documented", and in paraphrase those considerations include experimental design, data representativeness and construct validation. That supports a careful version of sentence 1: organizations managing AI risk are asked to think about whether test data represents real use. It says nothing about who licenses which records, so the sentence stays in the "test elsewhere" bin.
Sentence 3 belongs to SourceX's records, not yours. Under the terms, a company with any earlier application, relationship, lead, account or email correspondence with SourceX is a duplicate, and SourceX checks its own records before inviting an introduced company. A partner cannot see those records, so "never been asked" can only be tested company by company, as SourceX's duplicate decisions come back. Sentence 5 is not a claim any experiment can reach; it is a forecast, and it should never appear in a message to a founder.
| Story sentence | Kind of claim | Where it can be tested | What an introduction can observe | Result that would rewrite it |
|---|---|---|---|---|
| 1. Agents judged on real work | About AI developers | Research, standards and developer documentation | Nothing; a CEO cannot tell you what developers evaluate | Primary sources showing evaluation relies mainly on synthetic or public tasks in your target domain |
| 2. Records live inside operating companies | About companies | Referral pilot, metadata only | Whether a CEO describes dated records of work and outcomes, how many years, in which system, created by whom | Fewer than three of eight companies describe such records |
| 3. Never been asked | About SourceX's records | SourceX's duplicate review, one company at a time | Whether a company that applies is treated as new | Two or more of the companies that apply were already known to SourceX |
| 4. Our network reaches them | About the fund | Referral pilot | A named relationship owner who agrees to ask, then a CEO who wants the conversation | Three or fewer of eight have an owner willing to ask |
| 5. A category forms, we are early | About the future | No experiment can reach it | Nothing | Cut from founder messages; keep, if at all, as a labeled opinion |
Relationship and ownership evidence for each pilot company
A pilot company enters the experiment because it matches the story's archetype, never because the fund backed it. SourceX's screening does not consider funding: it looks at size at peak, years of records the company created itself, rights to license, a CEO who authorizes the discussion and fit with a current program. The fit facts are published: a peak of at least 50 full-time employees, contractors not counted, plus documented operating history, rights to license and location. 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 evidence comes in a fixed order, and none of it involves files, samples or logins. The relationship owner comes first, because the CEO decides whether the company discusses licensing and signs for it, and an introduction arranged without the decision maker knowing is usually not a fit. Material that belongs to the company's clients is usually not a fit either, which is why "who created the records" sits beside "how many years". No buyer relationship is inferred at any step: nothing in the pilot says or implies that any AI developer wants a particular company's records.
| Evidence | Who supplies it | When it is recorded |
|---|---|---|
| A named relationship owner who will ask | Your team, from the owner | Before any contact with the company |
| The CEO knows about the idea and wants the conversation | The owner, then the CEO | Before any introduction |
| Peak full-time headcount, contractors excluded | The CEO | First conversation |
| The year the company's own records start, and the system | The CEO | First conversation |
| Who created the records, and whether clients own any | The CEO | First conversation |
| Location, and the language of the records | The CEO | First conversation |
| Rights to license, relevance to current data needs and terms | SourceX's review | Only after the company applies |
An eight-company pilot with the losing results written down first
Choose eight companies the story is about, ask each relationship owner once, keep every question at metadata level and stop after three weeks. Write the predictions below before the first call and sign them, so nobody can move a threshold after the answers arrive. Eight companies cannot estimate a rate; they can break a confident sentence, which is all this pilot needs to do. For a fuller pre-registered design, the ten-introduction experiment sets out decision and stop rules.
The pilot ends early on two events: a relationship owner reports that asking harmed a relationship, or a CEO says the approach felt like pressure. Either one is a finding about sentence 4, and both outrank the schedule.
| Prediction (for sentences 2 to 4) | Measure | The sentence holds if | The sentence breaks if |
|---|---|---|---|
| Reach | Relationship owners willing to ask, of 8 | 5 or more | 3 or fewer |
| Consent | CEOs who want the conversation, of those asked | 3 or more | 1 or none |
| Records | CEOs who describe dated work-and-outcome records spanning 5 or more years, of 8 | 4 or more | 2 or fewer |
| Control | Companies whose main records belong to their clients, of those described | 1 or none | 3 or more |
| Newness | Applicants that SourceX's duplicate review shows it already knew | 1 or none | 2 or more |
What came back, and the story rewritten (Illustrative)
Illustrative results, invented for the method. Five relationship owners agreed to ask, and four of their CEOs wanted the conversation. Three of the eight companies described dated records of work and outcomes going back five years or more, which lands between the thresholds: neither holds nor breaks. One company's useful records turned out to belong mostly to its clients. Of the companies that applied, one had exchanged email with SourceX the year before and was treated as a duplicate.
Method note. The plan sorts story sentences by the kind of evidence each needs, using the program terms and screening guidance checked on 26 October 2026, NIST's AI RMF and the phi-1 paper read the same day. It measures reach, consent, record descriptions and duplicate outcomes at metadata level. It does not measure demand for any company's records, record quality, prices or payouts, none of which SourceX has published and none of which a pilot can observe. The revision log is the asset to keep: it shows the story shrinking toward what the fund can defend.
| Original sentence | Revised sentence | Evidence behind the revision |
|---|---|---|
| 1. Agents will be judged on real work | Risk frameworks ask evaluators to document whether test data represents real use; we have not tested what developers license | NIST AI RMF Map 2.3; no referral evidence either way |
| 2. The record of real work lives inside operating companies | Three of the eight established companies we asked keep dated records of work and outcomes going back five years or more | Pilot, metadata only; inconclusive against our thresholds |
| 3. Most have never been asked | At least one company we introduced was already in contact with SourceX | SourceX's duplicate decision |
| 4. Our network reaches their CEOs | Our relationship owners would ask five of eight; three CEOs are not ours to reach | Pilot reach and consent counts |
| 5. A category forms, and we are early | Cut | No experiment can test it |
Evidence against the story, and what would change the thesis
A story that survives its own pilot still faces evidence from outside it. Write the conditions that would change the thesis next to the story itself: if pilots in two different sectors both break on records or on reach, sentences 2 and 4 come out of the fund's narrative; if they hold twice, the story may claim what the revision log says and nothing more. Moving from a category bet to naming any company is a separate step with its own rules, covered in the difference between a category bet and a company endorsement, and a longer list of falsifiers is kept in what would disprove your network's data opportunity. Watching for shifts in data needs after the pilot is the subject of a platform team's early-warning system.
- Synthetic substitutes. The phi-1 paper describes a 1.3-billion-parameter code model trained on filtered web data plus about 1B tokens of synthetic textbooks and exercises generated with GPT-3.5, which "attains pass@1 accuracy 50.6% on HumanEval and 55.5% on MBPP". Where synthetic material can stand in for real work records, sentence 2 matters less.
- Demand sits outside the pilot. SourceX checks relevance to its current data needs in review, and criteria follow each program and can change. A sentence about records can be true while no current program wants them.
- Duplicates. Any earlier contact with SourceX makes a company a duplicate, so "never been asked" can fail quietly, one company at a time.
- Rights. Client-owned material is usually not a fit, and the most vivid records in a story are often the client's.
- Small samples. Eight companies can break a sentence; they cannot confirm a market.
- Relationship cost. Every ask spends an owner's goodwill, and no data on how owners or founders react has been published; record it in the pilot log.
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 pilot company has a willing relationship owner and a CEO who wants the conversation, the experiment can produce its first real introduction: register for a partner link and sign before submitting, check the company against who qualifies, and read the duplicate rules in the program terms before the introduction goes out. The fund-side overview is for venture platform teams, and related essays are in the guides hub.
Testing a category story through introductions is referral work; deciding whether SourceX itself is worth investing in is a separate diligence question, and investor materials, including the company and commercial metrics SourceX shares with investors directly, can be requested 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.