What a new data category means for your founder network: a signal map that turns AI research requirements into one question per founder, and the findings that should end the thesis

Short answer

A new data category does not make your portfolio eligible. It tells you which records to ask about, and so which founders to ask: those who worked at, sell to or buy from established companies that create those records every day. Translate the category's research requirements into signals a founder can see without touching records, ask each founder one question, and stop when the companies named turn out to hold their clients' records rather than their own.

Visual briefFull-size diagram
A three-column flow from what AI research needs to the company record that resembles it to who in a founder network knows that company; two rows lead to a founder question and the third stops because the records belong to clients.
Illustrative signal map. Rows follow the method on this page; no real companies or outcomes.

A platform team usually meets a "new category" as a paper, a benchmark or an announcement. The useful response is not a newsletter to the portfolio. It is a narrow network experiment with a written condition for giving up.

What the research says, and what it leaves out

Two recent publications show the kind of requirement a category brings. Neither names anyone as a buyer, and neither describes a SourceX program.

TypeStatementBasis
Research factOpenAI's GDPval evaluates models on real-world tasks across 44 occupations in the top nine sectors contributing to U.S. GDP, built from the work of professionals averaging about 14 years of experienceOpenAI's GDPval page, checked 2026-10-26
Research factIn agent evaluation, a task has defined inputs and success criteria, a transcript is the full record of a trial and the outcome is the final state of the environment, not what the agent says it didAnthropic's guide to evals for AI agents, checked 2026-10-26
Program factFunding is not among SourceX's screening checks; companies need 50+ full-time employees at peak, years of their own records, rights to license and an authorizing CEOWho-qualifies page and program terms, version 2026-10-22.2, since replaced by version 2026-11-04
ConditionalWhether records like these match a current SourceX program, and on what termsDecided only in SourceX review
Our thesisCompanies whose records resemble these requirements sit beside your founders, as former employers, customers and suppliers, more often than inside the portfolioA working hypothesis, not measured; the test below is how to check it

Anthropic's flight-booking example makes the point concrete: an agent can report a booked flight, but the evaluation checks whether a reservation exists. Requirements like that favor records where a second system confirms what happened. That is a property of a company's operations, not of its funding round, and it is what a founder can recognize from the outside.

Turning the category into a question a founder can answer

The category question, "who has agent evaluation data?", is unanswerable in a check-in. The company question is answerable: which established company you know closes work in a system that records the request, what was done and whether it worked, and has done so for years?

Founders know the answer for three kinds of company. Their former employers, where they saw the systems from inside. Their customers, whose operations their product touches. Their suppliers, whose service they receive. None of those companies is in your portfolio, and each has a CEO the founder may or may not be able to ask.

The category-to-company signal map

Each row starts with one requirement from the research, names the company record that resembles it, and says what a founder could observe without seeing a single file.

Research requirementCompany record that resembles itSignal a founder can see without recordsFounder position most likely to knowSignal that disconfirms the row
A task with defined inputs and success criteriaA service ticket or work order holding the request, the steps taken and a closure codeThe team closes work in one system that records whether the problem was fixedA founder who sells software to service or field operations teamsClosure codes are free text or missing, or the system belongs to the client being served
An outcome checked against the real end stateA claim paid or denied, a shipment delivered, an invoice matched to a paymentA second system confirms what happened: a ledger, a carrier scan, a payment fileA founder whose product connects to finance or logistics systemsOnly narrative notes exist, and nothing ties the work to an end state
Work like that professionals produce in their occupationEngineering drawings with revision history, audit workpapers, contract drafts with redlinesWork passes through a reviewer whose changes are keptA founder who came from a professional services firmThe deliverables belong to the firm's clients under their contracts
Grading calibrated against human expertsQA scorecards, inspection reports and second-review logs with the rubric in forceA QA team scores work against a written, versioned rubricA founder whose company supplies or buys from a contact center or manufacturerThe rubric changed often and old versions were not kept

Method note. Built on 2026-10-26 from the two publications above and SourceX's published screening profile. The map measures resemblance between a stated research requirement and a record type. It does not measure demand for any record, volume, price, or whether SourceX runs a program for the category; those are settled only in review. Every row is a hypothesis with its own disconfirming signal, and the disconfirming column matters more than the others. The evidence-gap note on outcome-linked demonstrations goes deeper on the second row.

The relationship and ownership evidence behind each named company

A founder naming a company is the start, not an introduction. Before anyone contacts a CEO, each named company needs five facts, gathered as metadata through the founder:

  1. Who the founder is to the CEO. Former colleague, vendor, customer. A vendor asking a customer's CEO spends commercial goodwill, and the founder decides whether that is worth it.
  2. Whether the founder will ask. A name without a willing asker stays on the list as "no path".
  3. Whose records they are. Records the company created itself, not files it holds for its own clients.
  4. The published fit facts. 50+ full-time employees at peak, contractors excluded; years of 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.
  5. The authorizer. The CEO decides whether the company discusses licensing, and signs for it.

Questions stay at the level of record type, rough count, years, system and administrator. No files, samples or logins move through the founder or the platform team.

Credit needs agreeing before the first call. A company yields credit for one partner only, whoever got a valid introduction in first during the 90-day attribution window, and nothing is credited before the company submits a verified application. If SourceX already holds any earlier lead, account or email exchange with the company, it is reviewed as a duplicate and no partner is credited. Whoever submits signs the SourceX Channel Partner Agreement first; if your platform signed it under the former title, Channel Partner Non-Circumvention and Non-Disclosure Agreement, that signed agreement remains in force. Decide in writing whether the founder registers or the platform team submits with the founder's consent.

A two-week founder round with its stopping rules written first

Test one row, not the whole map. Pick the row your founders are best placed to recognize and ask it inside check-ins you already hold, rather than in a separate campaign; the note on founder fatigue explains why.

StepWhat you doWhat you record
1. Choose the rowOne requirement, one record type, one signalThe row and the date
2. Ask twelve foundersThe company question, in their regular check-inFounders asked, founders who named anyone
3. Collect namesUp to three companies each, with the person who knows the CEOCompany, relationship, willing asker yes or no
4. Screen on metadataThe five facts above, through the founderFacts known, facts failing, whose records
5. Apply the rulesThe stopping rules below, written before step 2Keep, drop the row or drop the category

Write the decision rules before the first check-in:

  • Keep the row if at least three named companies pass the size and ownership facts on metadata and at least one founder will ask a CEO.
  • Drop the row if most named companies hold records that belong to their clients.
  • Drop the category for this network if founders name only venture-stage peers like themselves. That means your network does not reach the established operators the category depends on, and the founder alumni guide is the better next read.

Record every founder who named nobody. A round where eight of twelve founders draw a blank is a finding, not a failure to report. When a company passes, the founder asks the CEO; the platform team never contacts the company first.

Evidence that would weaken the category thesis

  • Supply can be commissioned. GDPval's tasks were built from the work of experienced professionals. Nothing on its page says they were licensed from those professionals' employers, and a buyer can recruit experts instead of licensing company archives. If the category's data keeps being made this way, it may never become a program for company records.
  • Outcomes are rarely recorded. Many companies log the request and the work but not a checkable end state. The second row fails more often than it passes.
  • Rights. SourceX's screening guidance treats material that belongs to a company's clients as usually not a fit, which removes many service providers from the third row.
  • Personal data. Categories that depend on health, biometric or other sensitive personal data face extra screening; the main site's guide to special category data in training datasets covers it.
  • Program fit. SourceX checks relevance against its current programs in review, and a research category is not a program. When criteria change, re-screen only the touched rows, as the criteria change register describes.
  • Founder goodwill. No data on how founders react to these asks has been published. Count declines, and stop the round early if founders signal the ask is unwelcome.

For the broader step from any thesis to a first introduction, see from AI data thesis to a useful company introduction; for the portfolio side of the program, the overview for venture platform teams.

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 founder names a company that passes the five facts and is willing to ask its CEO, register for a partner link, confirm the row against who qualifies and read the program terms before the founder makes the call.

Turning a research category into founder questions is referral work; judging SourceX itself as a possible investment is a different decision, for which the SourceX investor overview offers materials on request.

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

Does a new data category make our portfolio companies eligible?

No. SourceX screens on size at peak, years of self-created records, rights to license, an authorizing CEO and fit with a current program, and funding is not among the checks. A category only tells you which records to ask about and which founders might know companies that hold them.

Should founders forward the research to companies they know?

No. A forwarded paper reads as implied demand that nobody has confirmed. The founder who knows the CEO asks one plain question, explains the published fit facts and lets the CEO decide whether to talk to SourceX.

When should the signal map be redone?

Date every row. Re-run the rows a change touches when the research behind a category is revised or SourceX's program criteria change, and leave the other rows as they are.

Sources

  1. openai.com/index/gdpval
  2. anthropic.com/engineering/demystifying-evals-for-ai-agents
  3. refer.sourcex.si/who-qualifies
  4. refer.sourcex.si/terms
  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