Filtering is not bad. It just is not an advantage another network cannot copy by being more cautious. Learning compounds, because what one introduction teaches changes how the next ten are screened.
The bet, separated from the facts
| Type | Statement | Basis |
|---|---|---|
| Program fact | The partner dashboard shows each referral's stage and its reward status | SourceX how-it-works page, checked 2026-10-26 |
| Program fact | Criteria are set program by program, so a company turned down under one program may still be a fit for another one later | SourceX who-qualifies page |
| Program fact | An introduced company needs a peak of 50+ full-time employees excluding contractors, a documented history of operations, rights to license and an authorized contact | SourceX terms and who-qualifies page |
| Conditional | Whether any introduced company passes review or licenses records | SourceX review and a signed agreement |
| Thesis | Networks that record predictions and debriefs screen more accurately over successive cohorts than networks that only count introductions | Our hypothesis; no cohort results exist |
What a single introduction can teach
One introduction produces four kinds of lesson, and each can be written down in a sentence without naming the company.
| Lesson | Captured as | Recorded by | Example entry |
|---|---|---|---|
| Question quality | Which first question surfaced the fact that decided the case, or would have | The partner, after the stage changes | "Asking who administers the field-service system found an outside vendor holding the login" |
| Screening accuracy | A fit probability written before the introduction, compared with the stage reached | The partner, before and after | "Predicted 0.7; reached review" |
| Owner understanding | Whether the CEO raised a basic misunderstanding in the first follow-up, such as expecting to send data before any agreement | The partner, from the owner's own words | "Owner assumed customer files would be sent; corrected" |
| Handoff quality | Whether the introduction reached the CEO with consent and the metadata ready | The partner | "CEO knew in advance; record types and years ready" |
A decline teaches as much as a pass, sometimes more. The consultant network guide already recommends telling a partner quickly and kindly why an introduction was not a fit. This page turns that habit into something measurable.
Debrief without exposing the company
The debrief record holds categories, never content. Partners do not need to export, upload or describe confidential records, and nothing in the log should change that.
| Field | Allowed values |
|---|---|
| Case code | A random code, not a name or initials |
| Cohort | The quarter the introduction was submitted |
| Size band | Under 50, 50 to 199, 200 to 999, 1,000 or more full-time at peak |
| Sector and record family | Broad labels, such as "field services, work orders" |
| Prediction | A probability from 0 to 1, written before submission |
| Stage reached | As shown on the dashboard at the cohort cutoff |
| Deciding fact | Size, history, geography, rights, authorization, current demand, or not known |
| Earlier question | The question that would have surfaced the deciding fact, if any |
Two rules keep the log honest. Record "not known" when no reason is visible, instead of inferring one. And code current demand separately: SourceX checks relevance to its current data needs during review, so a decline on demand says little about the partner's screen. The rejection reasons study offers a fuller code list if your network wants one.
Change the questions between cohorts, never during one
A cohort is every introduction submitted in a fixed window, usually a quarter. Freeze the first-question list and the fit criteria for the whole window. At the boundary, change at most two items, and date the change in the log. Changing questions mid-cohort makes it impossible to tell which version produced which outcome.
A typical boundary change after a cohort with two agency declines on client-owned material: move "whose material is it, yours or your clients'?" from the fifth question to the first, and add "who administers the main system?" Nothing else changes until the next boundary.
The longitudinal learning-to-fit study
This is the asset: six measures recorded every cohort, with what learning looks like and the trap behind each one.
| Measure | Definition | Learning looks like | Trap |
|---|---|---|---|
| Fit rate | Introductions in the cohort that passed SourceX's initial review, divided by all introductions in the cohort | A rise that holds within size bands | Rises from filtering or a criteria change alone |
| Prediction score | Brier score: the mean squared gap between each prediction and the outcome, coded 1 for passed and 0 for not | A fall over cohorts | Open cases cannot be scored yet; report how many are pending |
| Unknowns at handoff | Average number of the four fit facts still unknown when the introduction was submitted | A fall | Partners may stop introducing anything uncertain |
| First-question yield | Share of debriefs where the deciding fact was one the partner's first questions covered | A rise | Depends on reasons being visible |
| Owner surprise | Share of introductions where the owner raised a basic misunderstanding afterwards | A fall | Partners may under-record their own misses |
| Mix | Introductions by size band and sector | Stable, or changes explained | A shifting mix makes every other measure hard to read |
Reading fit rates over time. Plot each cohort's fit rate on a p chart, the NIST handbook's chart for the fraction of units with a given outcome per sample, which treats each count as binomial with variance p(1-p)/n. With ten introductions per cohort, the limits are wide; a single good quarter falls well inside them. Only a run of cohorts above the center line, or a point outside the limits, says the rate has really moved.
Reading predictions. The Brier score, introduced in 1950 for weather forecasts, decomposes into reliability (do predictions of 0.7 come true about 70% of the time?), resolution (does the partner separate likely from unlikely cases?) and uncertainty (how hard the cohort was). A network that learns typically improves resolution first: it stops giving every company 0.5.
Minimum design. At least three cohorts, predictions recorded before submission, the same cohort cutoff age for every comparison, and mix reported beside every rate. The partner quality study compares roles at one point in time; this design follows one network through time, and the two can share a log.
Mandate changes and selection effects
Three things can masquerade as learning.
A program changed. The who-qualifies page says each program carries its own criteria, and that a company outside one program's criteria may come within another's later. If criteria or demand shift mid-study, fit rates move for reasons that have nothing to do with the network. Date every known change and compare prediction scores within the period before and after it, not across it.
The network got cautious. If introductions fall from twelve to four and the fit rate doubles, look at the mix column. A shift toward larger, older companies is filtering. Compare fit rates within size bands; if the within-band rates are flat, nothing was learned.
The first cohort was unlucky. A bad first quarter will usually be followed by a better one by chance alone. This is regression to the mean, and it is why one before-and-after pair proves nothing. Survivorship adds to it: partners who drop out after a bad cohort leave the pooled curve looking better than any partner's own.
Two relationships followed through (Illustrative)
Illustrative. Generic companies, invented facts, no outcomes implied.
A 180-person HVAC services company, a client of an outsourced controller in cohort 1. The controller predicted 0.7. The CEO agreed to the introduction and the metadata was ready: work orders in one field-service system for eleven years. The first follow-up revealed that an outside vendor administered the system, which slowed the next step. The case is still in review at the cutoff, so its prediction is not yet scored. Lesson logged: add "who administers the main system?" to the cohort 2 list.
A 120-person marketing agency, introduced by the same controller's network partner in cohort 1. Predicted 0.6. Not a fit: most of the material belonged to the agency's clients, which the who-qualifies page lists as usually not a fit. Brier contribution: (0.6 - 0)², or 0.36. Lesson logged: ask whose material it is first. In cohort 2, the network screened out two agencies before introduction and introduced one whose own project histories were company-owned. Whether that third agency passes is unknown; what is known is that the deciding question moved from the end of the conversation to the start.
The conditions for a durable learning advantage
- Predictions written before outcomes. Without them there is nothing to score, only stories told afterwards.
- Reasons, where they exist. A vendor-run survey for PartnerStack found 14% of 100 senior B2B SaaS leaders named lack of visibility into partner impact as their top obstacle to partner-driven revenue. Visibility gaps are common in partner programs; record "not known" rather than filling them in.
- Stable or dated criteria. A trend across an undated criteria change is not evidence.
- Enough volume. Below about ten resolved introductions per cohort, keep the log but hold the conclusion.
- Shared lessons. A lesson that stays with one partner is a habit. Shared across the network as an updated question list, it becomes the advantage.
- When to stop. If prediction scores stay flat for three cohorts on a stable mix, the network has learned what this program can teach it; keep the questions and stop tracking.
Faster learning leaves the fit facts unchanged
Learning changes how you screen, not what qualifies. 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. The company also needs 50+ full-time employees at peak, documented operating history and rights to license its own records. The CEO authorizes a licensing discussion and signs for the company, so every introduction reaches the CEO with their knowledge. Discovery stays at metadata level, using the columns in the metadata-only opportunity map; what stays unknown until SourceX reviews the company is whether its records match current demand and what any license would cover. Partners sign the SourceX Channel Partner Agreement, formerly titled the Channel Partner Non-Circumvention and Non-Disclosure Agreement, before submitting introductions. The four screening tests themselves are laid out in the network screening guide.
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.
Start the log with your next introduction rather than your next quarter: register for a partner link, screen the company against who qualifies, read the program terms, and write your fit prediction down before you submit.