# Evaluation vs training rights: clause checklists | SourceX

Canonical: https://refer.sourcex.si/compare/ai-evaluation-data-licensing-rights
Last updated: 2026-11-04

Training and evaluation licenses ask different scope questions because the records do different jobs. Training records change a model, so the questions are about which models, which methods and what survives the term. Evaluation records measure a model, so the questions are about staying out of training, who sees results and how long reruns are allowed. Each question below is tied to the clause of a signed license where counsel finds the answer, so a missing answer shows up as a missing clause.

These checklists sit under the broader comparison of an [evaluation-only data license](/compare/evaluation-only-data-license). That page explains why the two uses differ; this one is the working tool a company's counsel holds against an agreement.

## How to read a license with these checklists

Work from the signed or draft agreement, not from a summary of it. For each check, go to the clause named, and mark one of three results.

| Result | Meaning | What happens next |
|---|---|---|
| Answered | The clause states the answer in terms someone outside the deal could apply | Nothing; note the clause number |
| Silent | The clause exists but does not address the question | Counsel raises it with SourceX before signature |
| Conflicting | Two clauses point different ways | Counsel resolves which governs, usually through the definitions or an order-of-precedence clause |

Clause names vary between agreements. The five used here, definitions, grant, restrictions, term and termination, and ownership of outputs, are generic labels for where each topic usually lives.

## Checklist A: training permitted use

Training scope is about what the records may shape and what remains afterward.

| Check | Clause to read | Answered looks like | Flag if |
|---|---|---|---|
| Methods covered | Grant | Lists the permitted methods, such as pretraining or fine-tuning | "Any AI purpose" with no method or model named |
| Models and teams covered | Grant and definitions | Defines the licensee, its affiliates and contractors who may train | Affiliates undefined |
| Records in scope | Definitions | Names systems, date ranges and record types | Scope described only as "the dataset" |
| Exclusivity | Grant | States whether the training right is exclusive and for how long | Exclusivity implied but no term stated |
| Trained models after the term | Term and termination | Says what happens to models trained during the term | Return or deletion of records only, nothing on models |
| Redistribution | Restrictions | Bars resale or sublicensing of the records | Silent on onward transfer |

On exclusivity, SourceX licenses are exclusive for AI training; the term and scope are agreed in each deal. The trade-offs are set out in [exclusive vs non-exclusive data licenses](/compare/exclusive-vs-non-exclusive-data-license), and onward transfer in [whether licensed data can be resold](/questions/can-licensed-data-be-resold).

## Checklist B: evaluation permitted use

Evaluation scope is about separation. Google's machine learning course states the core rule: never train on test data, because duplicates across training and test sets mean the measurement stops reflecting how the model generalizes. The EU AI Act's Article 10 likewise treats testing data as its own category, used for an independent evaluation. A license that permits evaluation has to protect that separation in writing.

| Check | Clause to read | Answered looks like | Flag if |
|---|---|---|---|
| Held out from training | Restrictions | Bars any training use of the evaluation records, including by affiliates and contractors | Training barred for the licensee only |
| Leak handling | Restrictions or termination | Says what happens if evaluation items reach a training run | No consequence stated |
| Scores and quotation | Restrictions or confidentiality | States whether results may be published and whether records may be quoted | Silent on publication |
| Reruns on later models | Grant and term | Allows a defined number of reruns or a defined retention period | Retention "as needed" |
| Rubrics and labels | Ownership of outputs | Says who owns scores, rubrics or annotations built on the records | Silent on derived materials |
| Return or destruction | Term and termination | Certified destruction at term end, with a date | Destruction "on request" only |

The [rubric-based evaluation guide](/guides/rubric-based-evaluation-ai) shows what the derived materials in the fifth row usually look like.

## Where the checklists collide: one archive, split scope (Illustrative)

Illustrative and fictional. A 410-person third-party logistics provider has eight years of carrier dispute tickets. Leadership will consider training use for routine tickets and wants the 2,000 hardest disputes, those with a documented final ruling, kept for evaluation only.

| Collision point | Training answer | Evaluation answer | Clause that must reconcile them |
|---|---|---|---|
| Which tickets | Routine tickets, 2017 onward | The 2,000 ruled disputes, named by ticket ID range | Definitions: two separate defined sets |
| Overlap | Must exclude the evaluation set | Must never appear in training | Restrictions: express carve-out from the training grant |
| After the term | Trained models stay with the licensee | Copies destroyed after the last permitted rerun | Term and termination: separate survival rules |
| Shared | Customer contract limits and redaction | Customer contract limits and redaction | Warranties and the agreed redaction schedule |

If the definitions clause does not separate the two sets, every later clause inherits the ambiguity, which is why it is the first place counsel looks. None of this sets a price; terms are agreed in writing between the company and SourceX.

## What the partner records, and what only counsel reads

A partner's part is metadata: that the records exist, roughly how many and over which years, which system holds them, and who can authorize a licensing discussion. A partner never asks for samples and never reads the agreement. Gaps in permission are easier to catch before the first call; see [identifying missing permissions before a discovery call](/guides/how-to-identify-missing-permissions-before-a-discovery-call).

Introduced companies need 50+ full-time employees at peak (contractors excluded), documented operating history, rights to license and an authorized contact. 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. [Who qualifies](/who-qualifies) has the full list. What stays unknown until SourceX reviews the company: whether buyers want an evaluation set at all, whether the company's own customer contracts allow either use, and any offer.

## 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.

If an owner wants their counsel to test either scope against these checklists, [register as a partner](/register) before making the introduction, and point them to the [program terms](/terms) and [who qualifies](/who-qualifies); the reward is paid separately by SourceX and never comes out of the company's compensation.

## Sources

- https://developers.google.com/machine-learning/crash-course/overfitting/dividing-datasets
- https://eur-lex.europa.eu/legal-content/EN/TXT/PDF/?uri=OJ%3AL_202401689
- https://refer.sourcex.si/terms
