Most venture firms already have the raw material: partners circulate papers, model cards and memos every week. The platform team's job is not to summarize them for founders. It is to decide which ones justify asking a handful of CEOs a question, and to stop asking when the answers say the reading was wrong.
A reading list shows where partners look, not which companies hold records
A paper about how AI developers build evaluations says something about the field. It says nothing about whether a company in your network keeps records of its own, may license them or has a CEO who wants to talk. Those are the facts SourceX screens on: its screening page lists size at peak (50 or more full-time employees, contractors excluded), years of records the company created itself, the rights to license them, a CEO who authorizes the discussion and fit with a current program. Funding appears nowhere in that list.
So an agenda item has to cross a gap. It starts as a claim in a paper and has to become a question one CEO can answer about one company, without anyone outside that company seeing a file. And it has to stay a question. A reading list that turns into "our companies have exactly what the labs want" has skipped every step SourceX's review exists to take.
Five tests before a paper earns an agenda slot
A paper that fails any one test stays on the reading list and off the agenda. The tests are deliberately strict, because each admitted item spends a relationship owner's time.
- Primary. The paper, model documentation or regulator text itself, read in full by whoever proposes it. A newsletter summary does not count.
- About records a company makes. The claim implies something about the operating records companies create as they work, such as tickets, project files or case notes.
- Paired at admission. A credible counter-reading that predicts the opposite is named before the item is admitted, ideally by someone other than its champion.
- Testable in one quarter. The network can answer the question through conversations with CEOs it can already reach.
- No company data. The test runs on what CEOs say about their own records: type, years covered, system and who created them. Never files, samples or logins.
The research-to-action agenda template
Each admitted paper gets one row, and the agenda holds four rows a quarter. Every column is filled before anyone contacts a company; the last two change as the test runs.
| Column | What goes in it | Rule |
|---|---|---|
| Slot | One of the quarter's four | A fifth paper waits |
| Reading item | Full citation and the date it was read | Primary sources only |
| The paper's claim | One sentence in the paper's own terms, labeled as a paraphrase | No stretching beyond what the paper says |
| The team's inference | What the team thinks the claim implies about company records | Labeled as the team's, never as the paper's |
| Counter-reading | The source that predicts otherwise, and its claim | Required at admission |
| Network question | The question a CEO can answer about their own company | Answerable from metadata alone |
| Frame | Which companies, reached through which named relationship owners | Fixed before the first contact |
| Size and stop rule | How many conversations, and the count that retires the row | Fixed before the first contact |
| Owner | One person on the platform team | Not a committee |
| Review date | The last week of the quarter | No extensions without a new row |
| Status | Proposed, admitted, testing, kept or retired | Retired rows stay visible |
One row, filled in (Illustrative)
Illustrative. The two papers are real; the row, the companies and the numbers in the stop rule are invented to show the mechanics. The reading item is the SWE-bench paper, which describes a benchmark built from 2,294 tasks drawn from real GitHub issues and their pull requests across 12 popular Python repositories (paraphrased). The counter-reading is the Self-Instruct paper, which describes a model generating its own instructions, inputs and outputs, filtering the weak ones and being fine-tuned on the rest, with a reported 33% absolute gain for vanilla GPT-3 on Super-NaturalInstructions (paraphrased).
The pairing cuts two ways, which is why it earns a slot. SWE-bench shows evaluation tasks built from real, resolved work. It also shows them built from public repositories, and Self-Instruct shows training data a model wrote for itself; either could mean developers need fewer records from companies.
| Column | Entry |
|---|---|
| The paper's claim | A benchmark can be built from real issues and the changes that resolved them (paraphrased) |
| The team's inference | Companies with years of their own resolved work requests may hold material of a similar kind |
| Counter-reading | Public sources and model-generated data may cover the same need |
| Network question | Does your company keep its own multi-year records of resolved internal work requests, and in which system? |
| Frame | Twelve companies believed to have passed 50 full-time employees at peak, each with a named relationship owner, listed before contact |
| Size and stop rule | Twelve conversations; retire the row if fewer than 3 CEOs describe such records created by the company itself |
| Owner | Platform research lead |
| Status | Testing |
Writing the pilot protocol down before the first call
The row above is a small observational study, so it gets the same discipline as a large one. Four things are written down and dated before any relationship owner picks up the phone:
- Sampling frame. The twelve companies, by size band, sector and relationship owner. No replacements once contact begins; a company that cannot be reached stays in the twelve.
- Consent. The relationship owner asks the CEO first. The CEO agrees to a short conversation about their own records and can stop at any point. Nothing is noted beyond the variables below.
- Variables. Record type, years covered, system, whether the company or its clients created the records, whether the CEO would like a conversation with SourceX, and the date.
- Proposed analysis. Counts against all twelve: reached, answered, declined to talk, no reply, and records described. Descriptive only.
Declines and silences stay in the denominator
The row reports against all twelve companies, every time. A CEO who declines the conversation is a result, and so is one who never replies; deleting them would turn a test of the network into a test of the friendliest relationships.
Read the counts for what they are. If five of twelve CEOs describe their own multi-year work records, that describes this network this quarter, not the market. It does not show the paper was right about AI developers, and a CEO who wants to speak to SourceX has not been accepted by anyone. Retired rows stay on the agenda with the count that retired them, so next year's team can see which readings its network contradicted. For sorting disconfirming findings by how far they reach, use the disconfirming-evidence checklist; for turning a single thesis into a company question, the thesis-to-network canvas.
Method note. The five tests and the template's columns come from SourceX's published screening checks and attribution rule (the who-qualifies page and the terms, checked 2026-10-28); the worked row paraphrases the SWE-bench and Self-Instruct papers as posted on arXiv. The template tests whether a network can answer a question drawn from its reading. It does not measure demand for any company's records, SourceX's review decisions or any payout.
What the platform team may say to founders from the reading list
Sharing a paper with a founder is fine. Attaching a conclusion about their company is not. SourceX's screening page warns that "a newsletter line that reads 'our companies qualify' implies an acceptance nobody has given", and asks introducers to "Never list companies as eligible, and say 'may be a fit; SourceX decides after review'". The level of each thesis sentence, from category bet to company claim, is set out in the category-bet attribution guide.
| Say | Never say |
|---|---|
| "Our partners are reading this paper on how evaluations are built." | "This paper proves your data is valuable." |
| "Some companies may be a fit; SourceX decides after review." | "Our portfolio qualifies." |
| "If you want to talk to SourceX, that is your CEO's call." | "The fund recommends SourceX to you." |
| "We are asking a few CEOs one question this quarter." | "You would get priority access through us." |
| "Tell us about your records in your own words; we need nothing else." | "Send us an export so we can check." |
When the reading, not the network, should change
Four events reopen a row before its review date: the stop rule is hit; the counter-reading gains new primary support; SourceX's program pages change, since its screening page says each program sets its own criteria and a company that misses one may fit another later; or CEOs keep describing records that belong to their clients rather than themselves. A row that survives its quarter earns a second, wider test with a new frame, never a message to the whole portfolio. For a standing review rhythm, the early-warning cadence sorts signals by their distance from a SourceX decision, and the one-company conversation itself can follow the portfolio operations review agenda.
Where this agenda can mislead
- Reading what confirms. Teams read papers that flatter their bets. Requiring someone else to propose the counter-reading helps; it does not cure it.
- A weak counter-reading. A strawman paired with a favorite paper makes the gate decorative. The owner should be able to say what result the counter-reading predicts.
- Generated data. If methods like Self-Instruct keep improving, developers may need fewer records from companies, and rows built on real-record demand would retire one after another.
- Small rows. Twelve conversations describe twelve companies. They do not rank sectors or forecast anything.
- Relationship cost. Every conversation spends a relationship owner's goodwill, and nobody has published how CEOs react to repeated research questions from investors.
- Papers are not demand. No paper says what SourceX needs this quarter; only its program pages and its review do.
Program facts every agenda row inherits
A company that reaches the end of a row still has to fit on its own facts. 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 decision to open a licensing discussion, and the signature that follows, belong to the company's CEO. Whoever submits an introduction signs the Channel Partner Agreement first, and SourceX credits one partner per company from dated records once the company files a verified application inside the 90-day window; a link click does not settle it. Nobody on the platform team collects company records for SourceX at any stage.
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 row's conversations surface a CEO who wants to talk to SourceX, register as a partner, compare that company with who qualifies and read the program terms before anything is submitted; related thesis-to-network guides are in the guides hub.
Running a research agenda is referral work. Judging SourceX itself as an investment is another matter entirely, and the SourceX investor overview is where to ask for its materials.
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.