How to explain training versus evaluation briefly: two lines an owner can repeat, analogies for five kinds of company and one reply fixed before and after

Short answer

Say it in two lines, about the owner's own records. Training: an AI model learns from examples like your work orders, the way a new hire learns from past jobs. Evaluation: some of those records are kept out of training and used as a test, and the model is graded against what your team actually did. Then add the fact that keeps it honest: the license your CEO would sign after SourceX's review says which use applies, so you promise neither.

Visual briefFull-size diagram
One Illustrative stack of 14 years of work orders forks into two lanes: in the training lane the model learns from them and is changed; in the evaluation lane, kept out of training, the model works each case and is graded against the real outcome.
Illustrative. One company's records, two possible uses; the signed license says which one applies.

Owners who ask usually mean something else: will our records end up inside a model, is one use safer, and are you steering us? The card below answers all three.

Start from the records this company already keeps

"Your data" makes an owner picture customer lists and inboxes; "your 14 years of maintenance work orders" is something they can see and decide on. First get four facts at metadata level: the record type, the system, the years covered and how each outcome is written down. They are also your company-specific reason for raising the subject.

The outcome field matters most. SourceX's comparison of evaluation-only and training licenses notes that records showing real tasks with known outcomes, such as resolved tickets, are useful for both purposes, and that a license can name which one is permitted.

The two lines rest on definitions from outside SourceX. The EU AI Act defines training data as data used to train an AI system by fitting its learnable parameters, and testing data as data used for an independent evaluation of its expected performance. Google's machine learning course adds the rule that gives a test its value: evaluate on examples the model did not see in training. Anthropic's guide to agent evals describes the test itself: run the system on a task and let grading logic check the result. SourceX's referral site says it supplies licensed datasets for both uses.

The business-owner explanation card

Six slots, about 130 words. Lines 2 and 3 are the explanation; the rest make it about this company, keep it conditional and leave an easy choice.

SlotMost wordsWhat to say (fill the brackets)Rewrite it if
1. Why this company20"You have [years] of [record type] in [system], each closed with [outcome field]."You are guessing, or the records belong to the company's clients
2. Training25"Training: an AI model learns from examples like your [record type], the way a new [role] learns from past [jobs]."It treats "goes into the model" as settled
3. Evaluation25"Evaluation: some [record type] are kept out of training as a test; the model works each case and is graded against what your team actually did."It says the model never sees the records; working through them is the test
4. Who decides20"Which use applies is written into the license your CEO would sign, after SourceX's review."It promises a use, a price or acceptance
5. Your role25"I'd only make the introduction. If your company licensed records, SourceX would pay me a referral reward; it never comes out of your payment."The reward is missing, vague or buried after the request
6. The choice15"Want the link? If not, no reply needed."It asks for two actions or carries an invented deadline

Use one record noun in both lines: if line 2 says "work orders" and line 3 says "data", evaluation sounds like a bigger request. Keep the two sentences parallel, so "learns from" against "is graded on" carries the contrast. And translate technical words before they leave your mouth:

If you are about to saySay instead
Fine-tuning, pretrainingThe model learns from examples
Evals, benchmark, holdoutA test the model has not seen
Ground truth, labelsWhat your team actually did
Contamination, leakageTest cases slipping into training

Slot 5 is there because the FTC's endorsement FAQ treats a business relationship or payment as a material connection to disclose clearly, not vaguely. A private note to one owner is not necessarily advertising, but the plain sentence costs nothing.

Analogies for five kinds of company, and where each one breaks

An analogy from the owner's own business beats a definition. Each one breaks somewhere, and the break is where the next question lands.

Company (Illustrative)RecordsTraining is likeEvaluation is likeWhere it breaks
Property managerMaintenance work ordersA new coordinator reading past jobsRouting closed jobs they never saw, checked against what was doneA person forgets; what a model keeps after a license ends is a contract term
Corporate training providerCourse modules and quiz banksThe lessonsThe final examA test question that leaked into the lessons scores well and proves nothing
Engineering consultancyQA comments on drawingsAn apprentice studying marked-up drawingsA blind review of drawings with logged errorsA comment reversed later gives the test two answers
Software companyResolved support ticketsA new agent reading closed ticketsAnswering past tickets, scored against the recorded fixA polite reply can still be wrong; grading checks the outcome
Distributor's finance teamInvoice exception notesShadowing an experienced clerkAn audit sample checked against the recorded resolutionAn audit hunts for problems; a test set must also represent ordinary cases

The breaks come from the sources: Google's rule that a test case duplicated in training is not a fair test and that test sets must be representative, Anthropic's distinction between what an agent says and the outcome it produced, and the reversal question in the evaluation holdouts guide.

One reply to a COO, before and after (Illustrative)

Illustrative facts: a 230-person property management company, run by its founder as CEO, with 14 years of maintenance work orders in one system, each closed with a resolution code. The company's outside IT consultant knows the COO, who asks: "When you say AI training, what would they actually do with our work orders?"

Before: "Great question! Your data would be used for model training or evals. Evals are basically benchmarks, and eval data is really valuable right now and lower risk, because it never actually goes into the model. It's a great revenue opportunity, so it's worth moving quickly."

After: "Your 14 years of work orders, each closed with a resolution code, are the kind of record SourceX reviews for AI training and evaluation. Training means a model learns from them; evaluation means some are kept apart and the model is graded against what your team did. The license your CEO would sign after SourceX's review says which. I'd only make the introduction; if your company licensed records, SourceX would pay me a referral reward that never comes out of your payment. Want the link? If not, no reply needed."

Before saysThe problemAfter does
"training or evals", "benchmarks"Jargon explained with more jargonNames both uses in terms of work orders
"really valuable right now"A value claim nobody has made for these recordsLeaves price to SourceX's review
"never actually goes into the model"Wrong: in evaluation the model works through the recordsLeaves retention to the license
"great revenue opportunity", "moving quickly"An outcome and urgency with nothing behind themOne link, no deadline
No mention of the rewardThe partner's interest goes unstatedOne plain sentence before the request

The four questions owners ask after the two lines

Owner asksWhat you can sayWhat only the license or SourceX settles
"So our records end up inside the model?""Training: it learns from them. Evaluation: it works through them as test cases. Either way it uses them."What is kept, deleted or returned, and when
"Can we allow testing only?""You can raise it. SourceX treats evaluation-only as a narrower use to discuss, not a default."Whether it is available for these records
"Which one pays more?""Nobody has priced your records, and I won't guess."Price, agreed between the company and SourceX
"Who sees how the model scored?""That's a license term your counsel can check."Access to results and how long tests may be rerun

For the last question, the evaluation and training clause checklists show which clause answers each scope question. The third has a consequence for you: the program terms calculate the reward on the Actual Supplier Payout and do not distinguish training from evaluation, so you have no reason to steer.

Which records support either use, and who can say yes

Whichever use a buyer has in mind, an introduced company must reach 50 or more full-time employees at peak, not counting contractors, and bring documented operating history, rights to license the records and an authorized contact. Here that is the CEO, who authorizes a licensing discussion and signs; a COO can sponsor the idea internally. 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.

Personal consumer data and material that belongs to a company's clients are usually not a fit; drawings a consultancy made for a client may be the client's, which SourceX's rights review settles. Your questions stay at metadata level: what, how many, which years, which system, who administers it. No files, samples or logins. Demand fit, which use a license would name and what customer contracts permit stay unknown until SourceX reviews the company.

What the company signs, and what you earn under either use

The company signs its license directly with SourceX, the data is delivered and SourceX pays the company; your part ends at the introduction, as the guide to an introduction versus data delivery explains. You sign the SourceX Channel Partner Agreement before submitting any introduction.

The default reward on a company's first qualifying transaction is the greater of $50,000 plus 5% and 15% of Actual Supplier Payout, with no cap; Actual Supplier Payout is what SourceX actually pays the referred company. Further qualifying transactions with that company first agreed within 12 months of the first Closing pay 15%, and a transaction qualifies only from USD 100,000 unless SourceX agrees in writing. It is never calculated on SourceX's fees or revenue and never deducted from the company's payout. It falls due within 60 calendar days of whichever event is later, completed handoff or final written acceptance of the complete, validated dataset given by the Acquisition Partner, and the payment conditions in the terms apply. A later reduction of the supplier payout reduces the reward. Illustrative only: a USD 100,000 Actual Supplier Payout gives a USD 55,000 reward; USD 400,000 gives USD 70,000.

End with slot 6 and stop: your partner link, or the company fit checker for a non-binding look, with no deadline unless the company has a dated event of its own, such as a system being retired. Owners who want more can read the one-minute explanation or SourceX's page on training data versus evaluation data.

No test of these wordings with owners has been published. If your team adopts the card, log which slot draws follow-up questions, count them against every owner who heard it, and rewrite that slot first. More in the guides library.

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 an owner hears the two lines, understands that the license decides the use and asks how to start, sign up for a partner link, compare the company with who qualifies and keep the program terms open for the reward sentence.

  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

Should a partner suggest an evaluation-only license to a cautious owner?

No. Describe both uses accurately, say that evaluation-only is a narrower use the company can raise with SourceX, and leave the choice to the CEO and the company's counsel. The partner reward is calculated on the Actual Supplier Payout whichever use the license names, so there is nothing to steer toward.

Does evaluation mean the records never reach an AI model?

No. In evaluation the model works through the records as test cases and is graded on them; it does not learn from them. Who keeps the test set, who sees the scores and for how long are terms of the signed license.

Do I need a machine learning background to use the card?

No. The card needs the company's record type, system, years covered and how outcomes are recorded, which you can learn at metadata level. Technical questions about methods go to SourceX, and contract questions go to the company's counsel.

Sources

  1. refer.sourcex.si/terms
  2. refer.sourcex.si/who-qualifies
  3. refer.sourcex.si
  4. refer.sourcex.si/compare/evaluation-only-data-license
  5. eur-lex.europa.eu/eli/reg/2024/1689/oj/eng
  6. developers.google.com/machine-learning/crash-course/overfitting/di…
  7. anthropic.com/engineering/demystifying-evals-for-ai-agents
  8. ftc.gov/business-guidance/resources/ftcs-endorsement…

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