# Network experiment vs portfolio blast | SourceX

Canonical: https://refer.sourcex.si/guides/why-a-network-experiment-is-better-than-a-portfolio-blast
Last updated: 2026-11-04

Because a blast answers no question you can act on. Replies come from founders who were keen, not from companies that fit; nobody knows who read it, there is nothing to compare against, and silence is never recorded. Once every CEO has heard the pitch, you cannot run a cleaner test later. A network experiment writes your thesis down as a prediction, asks a few companies it ranks high and a few it ranks low through the people who know their CEOs, counts every decline, and says in advance which result would change the thesis. Nobody has run this design and published what it found.

This page is for VC platform leads and operating partners who hold a data thesis and a large network. It compares two ways of acting on the thesis and sets out one design. The other questions a launch raises are covered elsewhere: picking pilot companies and budgeting hours in [the ten-company pilot](/guides/choose-a-ten-company-pilot-before-a-portfolio-wide-launch), and the credit mechanics a mass email breaks in [activating a partner without a bulk email](/guides/how-to-activate-a-new-partner-without-a-bulk-email-blast).

## What a blast measures, and what it cannot

A portfolio-wide message measures attention. That is worth something for awareness and close to worthless for learning, for four reasons.

1. **No denominator.** You know who replied, not who read it, who forwarded it or who decided against it in silence.
2. **Self-selection.** The founders who answer are the ones with time and curiosity. Whether their companies fit is a separate fact the reply does not carry.
3. **No comparison.** Everyone got the same message at the same time, so nothing tells you whether the companies your thesis favors respond differently from the rest.
4. **A spent question.** After the blast, every CEO has heard the idea secondhand. A later, careful test runs on a contaminated list.

SourceX's screening page adds a fifth, which is about trust rather than method: "Screening comes before any portfolio-wide announcement, because a newsletter line that reads "our companies qualify" implies an acceptance nobody has given." Its guidance for a platform team that does write to the whole portfolio is to describe the criteria, invite CEOs to screen themselves, list no company as eligible and say SourceX decides after review.

## The targeted-learning versus broadcast comparison

| Criterion | Portfolio blast | Network experiment |
|---|---|---|
| Question it can answer | How many founders reacted to a message | Whether companies your thesis ranks high respond and fit more often than companies it ranks low |
| Denominator | Unknown readers | A frozen list of drawn companies, every one tracked to an outcome |
| Who speaks to the CEO | Nobody in particular; the message does | The person who already holds that CEO relationship |
| Comparison | None | Two strata, drawn by lot from your own list |
| Declines and silence | Invisible | Recorded with a reason code |
| Earlier contact with SourceX | Discovered after the fact, if at all | Asked before any submission |
| Reading as an endorsement | High: one sender, many recipients, an implied verdict | Low: one question, one CEO, no verdict |
| Cost to relationships | Spent on every founder at once | Spent on sixteen, by people who chose to ask |
| What you know after six weeks | Reply count | Counts by disposition in each stratum, against rules written first |
| Can you test again later? | Not cleanly | Yes; the rest of the list is untouched |

**Method note.** The criteria come from two places: what a design needs to support a conclusion (a known denominator, a comparison, recorded non-responses, rules fixed in advance) and SourceX's published rules (the CEO decides, credit goes to one partner per company, and earlier contact makes a company a duplicate under section 3 of the terms). The table compares designs, not outcomes. It does not measure demand for any company's records, record quality or payouts, which only SourceX's review and a signed agreement establish.

## A thesis worth testing, and the research that cuts both ways

Write the thesis as a claim about your own network that could be false. An example, labeled as our hypothesis and nothing more: *in our network, companies whose staff have written their own operating records for a decade or more will agree to a conversation and pass the published fit facts more often than our best-known growth companies.*

Research gives the thesis a reason and a limit. Muennighoff and colleagues, training over 400 models, report that with data constrained and compute fixed, "training with up to 4 epochs of repeated data yields negligible changes to loss compared to having unique data," and that with more repetition the value of added compute eventually decays to zero. Read one way, unique text matters once repetition stops paying, which favors companies holding long, self-written archives nobody has published. Read the other way, existing text can be reused several times at little cost, which pushes any scarcity later than a scarcity story assumes. The paper says nothing about any company, any buyer or SourceX's programs; it only explains why the thesis is worth testing and why it might fail.

Funding plays no part in either stratum. SourceX's screening page says none of its checks mentions funding: size at peak, records the company created, rights to license, a CEO who authorizes, and fit with a current program.

## Write the plan before anyone is asked

The Center for Open Science describes preregistration, in paraphrase, as recording hypotheses, methods and analysis plans before the research begins. A platform team can do the same in one dated page, shared with whoever will judge the result.

| Element | What you write down | Illustrative entry |
|---|---|---|
| Thesis as a prediction | The claim, and the direction you expect | "Thesis-high companies say yes with no failed fit fact more often than thesis-low" |
| Sampling frame | Every company in a defined segment that has a named relationship owner, list frozen on a date | 61 LP-owned companies and portfolio customers, frozen 3 November |
| Strata | Your ranking, made before any contact: high if you believe staff have written their own records for ten years or more, low otherwise; both must pass size on what you know | 24 high, 37 low |
| Draw | A fixed number by lot from each stratum | 8 and 8 |
| Consent | The relationship owner agrees to ask; the CEO agrees before anything is submitted; nothing is collected beyond what the CEO chooses to say | Owners who decline are recorded, not replaced |
| Variables | Asked or not, CEO reached, CEO's answer, each fit fact pass, fail or unknown, earlier contact with SourceX, whether the company applied | One row per drawn company |
| Analysis | Counts per disposition in each stratum, side by side; no percentages on groups this small | A two-column table |
| Decision rules | Set in advance (next section) | As written below |

Two consent points carry extra weight. Each ask goes through the person who already holds that CEO relationship, never from the platform team cold. And because section 2 of the terms credits only the first valid introduction of a new company, decide in writing before the draw who would submit for each company if its CEO says yes; whoever submits signs the Channel Partner Agreement first.

## Missing, declined and unsuccessful cases stay in the count

Every drawn company ends in exactly one of the first six dispositions, and all sixteen appear in the result; the last row is a later count on top.

| Disposition | Meaning | Counts toward the thesis? |
|---|---|---|
| Not asked | The relationship owner declined to ask | Reported, not a success |
| Not reached | The owner asked; no answer by the reading date | Reported, not a success |
| CEO declined | The CEO does not want the conversation | Reported, not a success |
| Fit fact fails | The CEO wants it, but a published fit fact fails on what the CEO said | Reported, not a success |
| Earlier contact | The company already applied to or exchanged email with SourceX | Reported separately; no partner is credited |
| CEO yes, no failed fact | The CEO wants the conversation and no known fit fact fails | The primary count |
| Applied | The company itself submitted an application | A lagging count, read when the 90-day window closes |

If you would rather score your judgment company by company, the forecast scoring in [the ten-introduction experiment](/guides/the-ten-introduction-experiment-for-an-ai-data-network-thesis) does that; the design here compares two groups instead.

Keep descriptive and causal claims apart. The draw is random within each stratum, so the sample is fair to your list. The strata are your own judgments, though, so a gap between them shows that your ranking predicts the outcome in this network. It does not show why, and it says nothing about networks you did not sample.

## Run it small, and say what would change the thesis

Six weeks of asks, one reading date, rules written before the first conversation:

- **Thesis supported, for this network:** the high stratum shows at least three more "CEO yes, no failed fact" outcomes than the low stratum, with at least four in total.
- **Thesis wrong, for this network:** the low stratum equals or beats the high stratum. Your ranking adds nothing; screen on the published fit facts alone.
- **Inconclusive:** anything else. Record it as a result, not a failure.
- **Stop early** if any relationship owner reports that an ask harmed a relationship, or if half the drawn companies turn out to have earlier contact with SourceX, which means the list is stale.

Treat any outcome as yours alone: no published result exists to compare it with. When a company in either stratum passes, the relationship owner asks the CEO, and the company applies and describes its own records.

## Where this comparison is wrong

- **A broadcast has a real job.** If the aim is to tell every CEO the criteria and let them screen themselves, a careful portfolio message in the form SourceX's screening page describes is the right tool. It just cannot test a thesis.
- **Sixteen companies detect only large gaps.** A modest real difference will read as inconclusive. That is why the rules demand a margin of three and a floor of four.
- **The research can turn.** The four-epoch finding already shows existing text stretching further than a scarcity story assumes; later work could stretch it further.
- **A right thesis can still produce nothing.** Rights and program fit are decided in SourceX's review, so companies that pass every fact you can see may still be declined.
- **Relationship cost is unmeasured.** No data on how owners or CEOs react to these asks has been published; the stop rule is the only guard.

For a checklist of findings that should end a thesis outright, see [what would disprove your network's data opportunity](/guides/what-would-disprove-your-network-s-data-opportunity); for weighing depth against reach in who you ask, [the specialist-network test](/guides/the-small-network-advantage-in-specialized-data-referrals); and for keeping any ask inside a founder's budget, [the founder fatigue guide](/guides/how-to-avoid-founder-fatigue-from-repeated-platform-offers). Turning a paper into the questions owners ask is covered in [using public AI research to ask better referral questions](/guides/how-to-use-public-ai-research-to-ask-better-referral-questions).

## 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 drawn company's CEO wants the conversation and no fit fact fails, the relationship owner can [register](/register); hold the CEO's answers up to [who qualifies](/who-qualifies) a final time and settle duplicates under the [program terms](/terms) before anything is submitted.

Designing a network experiment is referral work. Whether SourceX itself merits capital is a separate judgment, and SourceX takes requests for its investor materials through its [investor overview](https://sourcex.si/investors).

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.

## Sources

- https://refer.sourcex.si/who-qualifies
- https://refer.sourcex.si/terms
- https://www.cos.io/open-science
- https://arxiv.org/abs/2305.16264
- https://sourcex.si/investors
