The fit facts need no judgment: 50+ full-time employees at peak (contractors excluded), years of records the company created, rights to license them, a location and language that fit, and a CEO who authorizes the discussion. A row that fails one needs nobody's relationship. Whether to spend a founder's goodwill, whether the records are really the company's own, whether something you know came in confidence: that is the work a head of platform keeps.
What the founder gains, and what each lead hour costs
Founders gain from delegation only if it makes the asks better: a founder should hear about the program only when the company has already passed the published facts. A careless screener does the opposite, producing more asks, more of them irrelevant, each spending goodwill that belongs to whoever holds the relationship.
The lead's hour is the constraint. Spending it on rows that fail a published fact is waste; spending it on the rows below that need judgment is the job. The base case books referral income at $0, because SourceX has not published conversion rates or typical payouts and whether any company is accepted is decided only in its review. Delegation has to pay for itself in lead hours saved and founder asks improved, or not at all.
Decision rights: what the screener may decide alone
Give the screener this table on day one; anything not on it goes to the lead.
| Decision | Screener | Platform lead | Why it sits there |
|---|---|---|---|
| Record metadata: headcount at peak, years, systems, location, language share | Decides | Spot-checks | Published criteria; no relationship involved |
| Decline a row that fails a published fact | Decides, logging the failed fact and a revisit trigger | Samples weekly | The fact settles it |
| Count a staffing firm's employees | Decides: the firm's own full-time staff only, placed workers listed separately | Reviews | Settled program rule |
| Pass a row that clears every fact | Recommends | Decides whether to ask | Asking spends a relationship |
| Records held for customers or inside customer systems | Escalates | Decides | Client-owned material is usually not a fit |
| Group structures and earlier contact with SourceX | Escalates | Decides; SourceX settles credit in duplicate review | Credit questions the terms leave to review |
| Anything learned under confidentiality, such as board materials | Escalates | Decides | Confidential knowledge is a reason to ask the CEO, not content to pass on |
| Any contact with a founder or CEO | Never | Decides, through the relationship owner | Consent |
| Any mention of money or acceptance | Never | Decides the wording | The partner agreement bars implying a purchase or reward is assured |
Two rules from the partner agreement make the bottom rows non-negotiable. The fund may share SourceX's confidential information only with people who need it and are bound by similar duties, and it is responsible for their compliance, so the screener is acting for the fund. And the agreement gives partners no authority over a supplier's records, so a screener who is offered files has only one move: stop and escalate.
Ten training rows with the right call (Illustrative)
Illustrative. Generic companies, invented facts. Teach the reason, not just the call; the screener should be able to say which rule decided each row.
| Row | What the screener sees | Right call | The rule it teaches |
|---|---|---|---|
| 1 | 45 full-time employees, never more; nine years of service tickets | Decline; revisit if full-time headcount passes 50 | Size is judged at peak, and this peak fails |
| 2 | Peaked at 140 full-time employees in 2021, 38 today; twelve years of claim files | Pass to the lead | Peak counts, not today's number |
| 3 | Staffing firm with 70 internal full-time staff and 900 placed workers | Pass to the lead with both numbers | Only the firm's own full-time staff count; placed workers are listed separately |
| 4 | Distributor in Brazil whose records are about 60% English | Decline under the published language rule; note that program fit can change | Outside the U.S., Canada and the EU, at least 90% of the data must be in English |
| 5 | Software company whose valuable records sit inside customer tenants | Escalate: rights | Client-owned material is usually not a fit |
| 6 | Portfolio company in an active sale process, known from board materials | Escalate: confidential source and timing | Confidential knowledge is never screening material |
| 7 | Subsidiary whose sister company another partner introduced | Escalate: group credit | SourceX decides credit in duplicate review; one cap per group |
| 8 | Founder said "not this quarter" | Park; the lead decides when to revisit | Consent belongs to the company |
| 9 | A CFO emailed the screener a sample export | Escalate at once; do not open or forward it | Partners never handle records |
| 10 | 600-person logistics firm; the only contact is a functional vice president | Pass to the lead, marked "no relationship owner" | The CEO authorizes; the lead decides whether a path exists |
Rows 1 to 4 have published answers, so the screener decides them. Rows 5 to 9 each hit a named trigger. Row 10 clears the facts but has no path, which the lead solves or drops. Getting row 3 wrong means rereading the rules; getting row 9 wrong means a conversation before the next row.
The triggers, for the screener's wall: rights unclear; customer-owned systems; a group or an earlier contact with SourceX; information from a confidential source; any records seen or offered; any request to talk money; the relationship owner's hesitation. Each escalation carries the row, the fact, the trigger and where the fact came from, nothing more.
Calibrate before the screener works alone
Before the screener declines anything unreviewed, both of you classify the same 20 rows independently, using four calls: decline, pass, escalate, park. Then compare. The model comes from evaluation practice: Anthropic's guide to evaluating AI agents says automated graders should be calibrated closely against human experts so that the two rarely diverge, and it defines a transcript as the complete record of a trial, with its outputs, reasoning and intermediate steps, not just the final score. Here the screener is the grader, the lead is the expert, and the screener's written reason for each call is the transcript.
Overall agreement is not enough, because chance alone produces some of it. Cohen's kappa, as McHugh's review explains, compares observed agreement with the agreement expected by chance and runs from -1 to +1, where 0 means no better than chance and 1 means perfect agreement. A worked Illustrative round:
| Measure (Illustrative) | Value |
|---|---|
| Rows agreed, out of 20 | 18 |
| Observed agreement | 0.90 |
| Agreement expected by chance, from each person's call counts | 0.345 |
| Cohen's kappa | 0.85 |
| Trigger rows the screener escalated, out of 6 | 4 |
The lead called 8 declines, 6 passes and 6 escalations; the screener called 9 declines, 7 passes and 4 escalations, and both disagreements were rows the lead escalated. A kappa of 0.85 looks reassuring. The screener is still not ready, because the only two misses were exactly the rows that needed judgment. Make the gate two-part, fixed in advance: no missed trigger rows, then an agreement level you choose for the rest. Run a fresh 20-row set after any miss.
Work at each stage when no revenue is assumed
The planning figures are Illustrative, for one 25-company list; replace them with a time log after the first round.
| Stage | Screener | Platform lead | Referral income booked |
|---|---|---|---|
| Build and source the list | 2 hours | None | $0 |
| Metadata screen and decline log | 3 hours | 20 minutes sampling declines | $0 |
| Passes and escalations reviewed | 30 minutes preparing the rows | 1.5 hours deciding | $0 |
| Relationship-owner asks | None | 1.5 hours | $0 |
| CEO consent and introductions | None | Through the relationship owner | $0 |
| SourceX's questions about the introductions | Drafts answers | Approves them | $0 |
| Outcome back to founders | None | 1 hour | $0 |
In this illustration the lead spends a little over four hours on 25 companies, plus approving answers, instead of screening every row. Delegation earns its place only if those hours go to better asks.
Widen, pause or take the screen back
Write the thresholds down before the first round.
- Widen the screener's authority, so declines go out unreviewed and the lead samples one in five, after two consecutive calibration sets with zero missed trigger rows and agreement at or above your chosen level.
- Pause after any trigger row passed without escalation, any decline that failed no published fact, or a change in program criteria. SourceX judges each company against the criteria of the specific program it is being considered for, and today's decline can become a fit for another program later, so rules on the screener's wall go stale.
- Take the screen back if the screener contacted a founder or CEO without the lead's decision, opened or forwarded company records, or told anyone a company "qualifies".
A screen that should stop (Illustrative). An associate screens 48 companies in a week, passes 14 and emails four founders a template saying their companies "qualify". Every row was screened correctly on the facts. The delegation still fails: outreach happened without the lead's decision, and acceptance was implied to founders when only SourceX's review can decide it. The lead takes the screen back, writes to the four founders to correct the message, and restarts calibration.
Limits of a delegated screen
- It checks published facts, not demand. SourceX checks relevance to current data needs in its own review, so a perfect screen still produces declines later. The company fit checker is preliminary and non-binding for the same reason.
- Small calibration sets are noisy. Twenty rows teach a lot and prove little; treat one good round as a start.
- Delegation does not move confidentiality. The fund stays responsible for everyone acting for it.
- Turnover resets calibration. A new screener starts at row 1.
- No published measurements. SourceX has not published conversion rates or payouts, and time logs and founder feedback from a permissioned pilot have not been published; the hours above are placeholders.
Method note. The decision rights and training rows were built on 2026-10-26 from the program terms (Version 2026-10-22.2), the unsigned Channel Partner Agreement sample (version 2026-10-19), SourceX's who-qualifies page and owner decisions on staffing firms and parent companies, with calibration borrowed from Anthropic's evaluation guidance and McHugh's review of kappa. It does not measure screener accuracy in any real fund.
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 90% of the data is in English. For the screen itself, the portfolio screening workbook, the fit screening worksheet and the peak headcount decision tree give the screener its forms; the overview for venture platform teams covers the program from the fund's side.
Next step
Refer a company. Earn up to $100,000. Default reward: 25% of Actual Supplier Payout, capped at $100,000 cumulatively per referred company. Eligibility and payment conditions apply, and a signed Schedule A or addendum can differ. A lead, meeting or signature alone does not trigger payment.
Run the 20-row calibration before you hand over a single list, then register for a partner link, give the screener the criteria on who qualifies and keep the program terms with the lead, who makes every ask.
A fund weighing SourceX as an investment, rather than as a program to run, should treat that as its own diligence question; the SourceX investor overview takes requests for investor 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.