When to stop a referral experiment early: harm, capacity and futility rules a platform team fixes before the first ask, checks on set dates and never triggers because one version is ahead

Short answer

Stop a referral experiment early for one of three reasons, each written into the plan before any CEO is asked. Harm: someone is misled, pressured or exposed, so asking stops on the first report. Capacity: your team can no longer run the protocol as written, so new asks pause in both versions until it can. Futility: counting shows the remaining companies cannot meet your decision rule, so the test ends and is reported. An early lead is never a reason to stop; it waits for the reading date.

Visual briefFull-size diagram
Left, three stop cards (harm stops on one report, capacity pauses both versions, futility ends the test when the rule is out of reach) above a struck-through card saying an early lead never stops a test. Right, an Illustrative look 2 with twelve slots per version: A has 5 agreed, B has 2 agreed and 3 slots left, so B's best case only ties A and the test stops as futile.
Illustrative. Invented counts that follow the worked example on this page; no SourceX result is shown or implied.

This page is a design: no results sit behind it. It covers what each scheduled look between the first ask and the reading date may decide. The market-by-market pilot shows how to pause one office; this page supplies the timing and arithmetic behind any early stop.

Three reasons to end early, and the one that never counts

Each legitimate reason protects something. A harm rule protects people: the CEO, the relationship owner, the company's confidential information. A capacity rule protects the protocol, because a test run differently from its plan answers a different question. A futility rule protects attention: asking more CEOs once the answer cannot change wastes their time.

An early lead protects nothing. It is also the reason most likely to be wrong. With two dozen companies, the version that leads after six asks often stops leading after twelve, and a team that ended the test at six never finds out.

The ICH E9 guideline on statistical principles for clinical trials is direct about when the plan for interim looks must exist: "This necessitates that the protocol (or appropriate amendments prior to a first analysis) contains statistical plans for the interim analysis to prevent certain types of bias." Trials that stop for benefit do so against statistical boundaries planned in advance. A referral test of 24 companies is far too small for such a boundary to carry weight, so this design drops the option entirely. CONSORT's explanation of its item on why a trial ended adds the reporting angle: readers can infer less from a trial cut short on the strength of its own data than from one that ran to a goal set independently of results (our paraphrase).

The harm-capacity-futility stop rules

Copy it into the plan, set the thresholds and date it before the first ask.

RuleFires whenEvidence neededWhat happensRestart only when
HarmA CEO says the approach felt like pressure; a message implied the company qualifies or that SourceX endorses the fund; anyone offers or asks for records, exports, samples or logins; a relationship owner reports a strained relationship; a fact learned through a board seat travels beyond one question to that company's own CEOOne report, from anyone in the testAll new asks stop the same day, in both versions; replies to people already in conversation continueThe cause is named and fixed, the person affected has had an answer, and the look owner signs a dated restart note
CapacityFollow-ups slip past the window each CEO chose; the dated record falls more than a week behind; a CEO's question waits longer than the protocol promised; nobody checks a name against the cohort list before an askTwo missed promises in a fortnight, or one gap in the recordNew asks pause in both versions; nothing already promised is droppedThe backlog is cleared; a pause that passes four weeks becomes a stop
FutilityEven if every remaining company went the new version's way, the decision rule could not be metThe counts at a scheduled lookThe test ends; open conversations run to their natural endNot in this cycle; a fresh plan is written
Early leadOne version is ahead at a lookNone, because it never firesNothing; the test continues to the reading dateNot applicable

Harm and capacity rules both act on both versions. Until a harm's cause is known, nobody can tell whether the version or the process produced it; a faulty version is withdrawn and the comparison ends, while a faulty process is fixed and both versions restart together. Pausing only a strained version would put the two in different weeks, so anything that shifts over time, a holiday stretch or a news story, would land on one alone.

Method note. The table applies three principles from clinical trial practice (interim plans fixed in advance, safety reviewed continuously, stops reported with their reason, from ICH E9, FDA's guidance on data monitoring committees and CONSORT 2025) to triggers drawn from SourceX's terms, who-qualifies page and program rules, all read on 2026-10-28. The thresholds (one report, two missed promises, four weeks) are suggested defaults, not measured values. The table measures nothing about demand, record quality or payouts.

Fix the looks and the look owner before the first ask

Three kinds of look, three clocks:

  • Harm watch, continuous. Every report reaches the look owner the day it arrives; a harm rule should never wait for a meeting.
  • Capacity check, every two weeks. Read from the dated record: promises made to CEOs, promises kept, questions open.
  • Futility look, at fixed counts of CEOs asked. For example, at the halfway and three-quarter marks, so the look cannot drift to whenever the numbers seem interesting.

The look owner is one person who makes no asks and has no stake in the new version. FDA's guidance on data monitoring committees calls helping to ensure subject safety a prominent responsibility of the committee, and frames futility monitoring as asking whether the trial still has a reasonable likelihood of reaching its objective (our paraphrase). A referral test needs one person, not a committee, but the principle holds: whoever wants the new version to win does not decide when the test ends.

A look may not change the measure, move the decision rule, add companies to rescue a version or push back the reading date. Each of those is an amendment. Date it, report it, and label everything after it exploratory.

The futility check, worked through (Illustrative)

Illustrative. An invented platform team and invented counts; no SourceX result is shown or implied.

A platform team names 24 companies that pass the published fit facts on metadata alone, and draws 12 for each version before any outreach. Version A is the current written explainer. Version B is the same explainer plus an offer of a 15-minute call with the platform lead before any introduction. The measure is CEO-agreed introductions: a CEO's dated written yes to being introduced. The decision rule, fixed before the first ask, reads: adopt B only if, at the reading date, B has at least 3 more CEO-agreed introductions than A; otherwise keep A. Futility looks fall after 6 and after 9 asks per version.

The check is one subtraction. B's best possible final count is B's count so far plus B's asks remaining. A's lowest possible final count is A's count so far. If the first minus the second falls below the margin of 3, B can no longer meet the rule.

LookAsked per versionA agreed so farB agreed so farAsks left per versionB best case minus A lowestMargin neededVerdict
163268 - 3 = 53Continue
295235 - 5 = 03Stop: futile

At look 2, even if all three remaining B companies said yes and no remaining A company did, B would only tie A. The test ends, A stays, and the report says B was not shown to be better, a weaker claim than "B is worse", which nine companies a side cannot support.

The check never condemns A, since the rule only asks whether B wins, and never fires in B's favor: had B led 5 to 2 at look 2, the test would have run on to twelve a side. A one-version pilot uses the same check. If the plan needs 6 CEO-agreed introductions from 20 asks and 16 asks have produced 1, the most it can reach is 5, and it stops. Why a short lead proves so little is set out in how to avoid declaring a winner from five replies.

The stop note: what to write the day a rule fires

CONSORT 2025 gives the reporting item a plain name: "If relevant, why the trial ended or was stopped." A referral test's stop note answers it in six lines:

  1. The rule that fired, quoted from the dated plan, and the date it fired.
  2. Every company in the cohort at its furthest dated stage: named, asked, declined, silent, agreed. Include the companies this test will now never ask.
  3. Who called the stop and what evidence they saw.
  4. What happens to open conversations. A stop ends new asks; it never takes back an introduction a CEO asked for, and whether that company continues with SourceX stays its CEO's decision.
  5. What the counts can and cannot say: descriptive only, no causal claim beyond the decision rule.
  6. What did not work, given the same space as what did.

A stop leaves credit untouched. Section 2 of the terms credits "the first valid introduction of a new company within the 90-day attribution window" once that company submits a verified application, decided from dated records, and adds that "an earlier link click alone does not decide credit." For the write-up that goes to partners and LPs, follow the pilot retrospective outline. Harm that surfaces after the stop belongs in the side-effect review.

Which companies may enter the cohort

Screen before any company is drawn into a version, from conversations with whoever owns the relationship rather than from documents. Each company needs at least 50 people on its full-time payroll at its largest, with contractors left out of the count (a staffing firm counts only its own staff, never workers placed with clients), an archive of operating records it produced and still controls, the legal right to license that archive, and a CEO who has heard the idea and wants to talk. 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.

Three things stay unknown until SourceX reviews a company: whether its records match a current program, whether customer contracts limit their use, and what a license would cover. Nobody in the test gathers records, system lists or headcount files for SourceX. If a cohort company is acquired mid-test, its CEO decides until closing and the acquirer's CEO after. Whoever submits introductions signs the Channel Partner Agreement first. A stop rule governs how your team asks, never whether a company fits.

Where early stopping misleads

  • The margin is a judgment. The arithmetic is exact; the margin it tests is a choice. Set it too high and almost every new version looks futile by the second look, so record why you chose it.
  • A harm stop splits the sample. Companies asked after a fix met a different protocol. Report the before and after groups separately.
  • Frequent looks invite amendments. Each look at the counts adds pull to adjust something, which is why ICH E9 wants interim plans fixed before any analysis.
  • The measure stops before SourceX. CEO-agreed introductions are visible to your team; SourceX's review and any completed handoff come later, on timing SourceX does not publish, so no futility look can rest on completed transactions. The learning review for pilots with no completions covers that gap.
  • Some tests should not run at all. If withholding the new version from half the cohort would itself be unfair, the holdout decision tree names a design that holds nobody back.

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.

Before the first ask, put the stop table and look dates in the dated plan, register as a partner under the entity that will submit, screen every cohort company against who qualifies and check the program terms version your plan relies on. More test designs sit in the partner guides.

Deciding how to run a referral test is a network question. Evaluating SourceX as a company to invest in is a separate one: its investor page takes requests for investor materials, and the company says it gives investors company and commercial metrics directly.

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

Can a platform team stop early if the new version is clearly ahead?

Not under this design. With a dozen companies per version, a lead at a scheduled look is the result most likely to shrink by the reading date, and stopping on it removes the chance to see that happen. Let the test run; the decision rule settles it at the end.

Does stopping a test cancel introductions a CEO already asked for?

No. A stop ends new asks and nothing else. A CEO who asked to be introduced stays introduced, the company decides for itself whether to go on with SourceX, and credit follows the published terms: the first valid introduction inside the 90-day attribution window, once the company submits a verified application.

Who should own the stop decision on a small platform team?

One person who makes no asks and has no stake in the new version, such as a partner outside the platform team or the fund's operations lead. They receive every harm report the day it arrives and run the scheduled capacity and futility checks from the dated record.

Sources

  1. refer.sourcex.si/terms
  2. refer.sourcex.si/who-qualifies
  3. database.ich.org/sites/default/files/E9_Guideline.pdf
  4. fda.gov/media/176107/download
  5. consort-spirit.org/item23b-whythetrialended
  6. 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