# Referral experiment causal claims checklist | SourceX

Canonical: https://refer.sourcex.si/guides/how-to-avoid-calling-correlation-a-referral-program-success
Last updated: 2026-11-04

Treat every success sentence as a claim that has to pass review before it reaches an LP letter, a board pack or a partner newsletter. Write down the comparison the sentence rests on, test it against the rival explanations that raise introductions without any help from your program, and let whatever survives choose the verb. A rise after a founder session "followed" it. Only a comparison fixed before the period began, with every company counted, earns "caused".

This page reviews claims once the counting is done. Whether to hold a group back belongs to [when a pilot needs a holdout group](/guides/when-a-pilot-needs-a-holdout-group), and how lopsided a small split must be before it means anything belongs to [declaring a winner from five replies](/guides/how-to-avoid-declaring-a-winner-from-five-replies). No program results appear here, SourceX's or anyone else's.

## Credit tells you who, not whether

SourceX's terms settle a question that sounds causal and is not. Section 2 reads: "The first valid introduction of a new company within the 90-day attribution window receives credit when that company submits a verified application," and attribution "is decided from dated records". That rule says who receives credit. It is silent on whether the company would have applied anyway, through someone else or on its own schedule. Section 3 opens a second gap: a company that already had an application, lead, account or email correspondence with SourceX is a duplicate and earns no credit, so some of the companies your network knows best never show up in a credited count, whatever your program did.

Statisticians draw the same line in general terms. The Australian Bureau of Statistics treats correlation as a measure of how large a relationship between variables is and which way it runs, and keeps causation for the case where "one event is the result of the occurrence of the other event". Its example: "smoking is correlated with alcoholism, but it does not cause alcoholism." HM Treasury's Magenta Book, the UK government's evaluation guidance, judges impact against a counterfactual, a group standing in for what would have happened without the intervention, and counts a before-and-after reading of the treated group alone as a weak stand-in (paraphrase).

## Three kinds of sentence, and what each one needs

| Sentence type | Example (Illustrative) | Evidence it needs | Verbs it may use |
|---|---|---|---|
| Descriptive | "From March to June we raised the program with 22 CEOs; 5 agreed to an introduction and 2 companies reached a verified application" | Every company asked, its last recorded stage and the date of the count | asked, agreed, reached, recorded |
| Associational | "CEO-agreed introductions were more common after the founder session than before it" | A stated comparison, the rivals that could explain it, and both sides counted the same way | followed, coincided with, was higher among |
| Causal | "The founder session increased CEO-agreed introductions" | A comparison group fixed before the period, the same measure in both groups, every company counted and each rival ruled out | caused, increased, produced, drove |

A sentence may always step down a row. It may never step up one without the evidence the higher row asks for. "Our program works" is a causal sentence in casual clothes, and "founders love it" is descriptive only when you can show how many founders were asked.

## The causal-claim review checklist

Run it in writing on the exact sentence you want to publish, before anyone outside the team reads it. Each check has a pass condition and a consequence.

| # | Check | Passes when | If it fails |
|---|---|---|---|
| 1 | The claim is written out | One sentence names the unit (a company, not an email or a partner), the outcome and the period | Nothing can be reviewed yet |
| 2 | The outcome is the planned one | The measure was named before counting began, such as CEO-agreed introductions that reached a verified application | Descriptive only, marked as chosen afterward |
| 3 | The comparison came first | A dated plan names the comparison group or period | Associational at most |
| 4 | Every company asked is counted | Declines, silences, companies that did not fit and companies still open all appear at their last stage | Fix the count before any sentence |
| 5 | Timing lag | Each outcome is matched to the ask that produced it; a credit this quarter can come from an ask last quarter, inside the 90-day window | Re-date outcomes to asks, or drop the period comparison |
| 6 | Selection | Both groups were chosen the same way, and the new approach was not tried only on the warmest relationships | Associational at most, with the selection stated |
| 7 | Would have happened anyway | A comparison shows the background rate, and companies already in contact with SourceX are left out of both groups | Associational at most |
| 8 | Something else changed | No new terms version, program change or other campaign fell inside the period, or it reached both groups alike | Name the change beside the claim |
| 9 | Rebound from a slow spell | The baseline was not an unusually quiet quarter | Compare with several earlier periods, or say so |
| 10 | One person carried it | The result holds without the single most active relationship owner | Credit the person, or report both readings |
| 11 | Small numbers | The split clears a threshold written down before the count | "No difference we can tell apart from chance" |
| 12 | No company can be identified | No sentence, table or chart lets a reader name a company or its outcome | Remove the detail |

**Method note.** Checks 5, 7 and 8 rest on SourceX's published rules, read on 2026-10-28: the 90-day attribution window and dated records (terms section 2), the duplicate rule (section 3), dated terms versions (section 11) and the who-qualifies page's statement that each program applies its own criteria and that criteria can move. The other checks restate the general logic of the ABS and Magenta Book pages cited here. The checklist has not been tested on any network's data, and it cannot tell you whether your outcome measure was worth counting.

## Hypothesis, unit, measure, stop rules and the decision rule

A claim review is a small design in its own right, so write it down before the reporting period ends.

- **Hypothesis under review.** The exact sentence, such as "the founder session increased CEO-agreed introductions".
- **Unit.** The company whose CEO was asked. Partners, emails and events are never units.
- **Measure.** The outcome named in the plan, counted at a fixed number of days after each ask.
- **Stop rules.** Withhold the claim and stop the review if the comparison was picked after the numbers were seen and no other comparison exists, if any company asked is missing from the count, or if a sentence would let a reader identify a company. Pause without withholding when SourceX published a new terms version or a program changed its criteria during the period: record the change, then run checks 7 and 8 again.
- **Decision rule.** Failing check 1 or 4 means no claim until it is fixed. Failing check 2 or 12 leaves a descriptive sentence or nothing. Failing check 3 or any of checks 5 to 11 caps the claim at associational, with the failed check named next to it. A causal verb needs every check to pass.

## Three claims through the review (Illustrative)

Illustrative. Invented networks and counts; no real program or outcome.

**Claim A: "Our founder session doubled introductions."** Founders chose whether to attend. Of 11 who came, 4 later agreed to an introduction; of 19 who stayed away, 2 did. Check 6 fails, because founders who turn up to a session on the subject were already more curious, and check 11 fails at these counts. Published as: "Of 30 founders invited, the 11 who attended agreed to introductions more often than the 19 who did not. Attendance was voluntary, so the session cannot be separated from founders' earlier interest."

**Claim B: "Since we rewrote our introduction email, verified applications are up."** The new email went out in May. That month SourceX also published a new terms version, and the network's largest LP began mentioning the program to its own companies. Checks 3 and 8 fail. Published as: "Verified applications from May to August were higher than from January to April. The email changed in May, and two other changes fell in the same month, so this report cannot say which mattered."

**Claim C: "Version B of our message produced more CEO-agreed introductions."** Allocation was a coin toss per company, written down before the first send, and every company is counted at day 90. Version B: 5 of 14. Version A: 3 of 14. Checks 1 to 10 pass; check 11 fails, because a split of five to three is well inside what chance alone produces. Published as: "In a randomized comparison of two truthful messages, 5 of 14 and 3 of 14 CEOs agreed to an introduction. The gap is too small to call, so the network keeps its current message and runs the comparison again."

## Report what did not work, and keep description apart from cause

A report that carries only the sentences that passed hides the ones that failed. Put three lists in every write-up: claims published at full strength, claims downgraded and the check that downgraded them, and claims withdrawn. Keep the descriptive tables, every company at its last stage, in a section of their own, so a reader can check the counts without accepting your interpretation of them.

Most pilots end before anything completes, because a payable reward sits behind a long chain of conditions in terms section 5. Report open companies as open; [what to do when a pilot has no completed transactions](/guides/what-to-do-when-a-pilot-has-no-completed-transactions) sets out that review. For the structure of the whole write-up, see [a pilot retrospective investors can trust](/guides/how-to-publish-a-pilot-retrospective-investors-can-trust), and for a design built around a single switch date, [what changed after we explained eligibility up front](/guides/what-changed-after-we-explained-eligibility-up-front).

## Where the review itself can mislead

- **Too strict to learn from.** A team that publishes only causal sentences will publish almost nothing. Descriptive counts with full denominators are worth sharing on their own.
- **A comparison is not always possible.** Holding companies back can be unfair or impractical, and the holdout guide says when not to try. Without a comparison, the strongest honest verb is associational.
- **The outcome may be the wrong one.** A clean causal claim about replies is still a claim about replies. Whether a company is accepted, what it licenses and whether any reward follows are decided in SourceX's review, often after a pilot has closed.
- **Checklists invite box-ticking.** Someone who did not run the program should complete it.
- **Programs move.** Criteria follow each SourceX program and can change, so a comparison that spans a change in criteria compares two programs, not two approaches.

## 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 claim survives review and the companies behind it have CEOs who want the conversation, [register as a partner](/register), recheck each company against [who qualifies](/who-qualifies) and read the [program terms](/terms) before anyone is introduced.

Testing claims about your network is referral work; judging SourceX as a possible investment is a separate exercise, and investor materials can be requested through the [SourceX 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/terms
- https://refer.sourcex.si/who-qualifies
- https://www.abs.gov.au/statistics/understanding-statistics/statistical-terms-and-concepts/correlation-and-causation
- https://www.gov.uk/government/publications/the-magenta-book/magenta-book-central-government-guidance-on-evaluation-html
- https://sourcex.si/investors
