This analysis is part of the reinforcement learning with verifiable rewards guide. The qualification scorecard decides whether to raise the topic at all; this page is for describing what you found.
What a verifiable-rewards review expects
A verifiable reward needs a check a program can run. The Tulu 3 paper defines RLVR as rewarding a model only when a verification function confirms its output is correct, and DeepSeek-R1 makes the same requirement concrete: answers must arrive in a specified format so a rule can verify them. For a business record, the equivalent is a complete set of evidence around one task, so a check can compare what was done with what should have been done.
| Evidence | What it answers | Illustrative example: inventory write-down |
|---|---|---|
| Task | What was asked, by whom, by when | Close item: review slow-moving stock |
| Inputs | What the person could see at the time | Aging report as of the period end |
| Rule version | Which policy applied on that date | Reserve policy v4: 50% after 365 days without movement |
| Approver identity | Who accepted it, separate from who did it | Named controller, timestamp |
| Action log | What was actually done | Reserve entry with amount and item list reference |
| Final outcome | Whether the result stood | No reversal; later disposal matched the reserve |
What companies typically have, and what is usually missing
Illustrative pattern drawn from how common systems are used, not survey data. Proportions vary widely between companies.
| Evidence | What companies typically have | What is usually missing | Effect on the check |
|---|---|---|---|
| Task | Close checklists, tickets, work orders | Due dates and owners in older years | Minor: the task can often be reconstructed |
| Inputs | The final report or document | The version as it stood on the day | The check may compare against later data |
| Rule version | The current policy | Earlier versions with effective dates | Cannot tell which rule applied |
| Approver identity | An approval flag or status | A named person distinct from the preparer | Cannot test segregation of duties |
| Action log | The posted entry or transaction | Edits made before approval | Minor for outcome checks, larger for process checks |
| Final outcome | The original record | The link to later reversals, credits or disposals | Cannot tell whether the result stood |
Rule version, approver identity and final outcome matter most, because no amount of volume compensates for them. A verification function is only as good as the evidence it checks: an archive without the final outcome can confirm that a task looks complete, not that it was right.
How to describe each gap honestly in an introduction
Use one sentence per gap, stating what exists, what does not and who could confirm it. Avoid both minimizing ("minor data issues") and dramatizing ("records are incomplete").
| Gap | Honest wording | Avoid |
|---|---|---|
| Rule version | "Current policies are kept; earlier versions before 2022 were overwritten." | "Policies are documented." |
| Approver identity | "Approvals are recorded by status; names were added when the new ERP went live in 2021." | "Full approval trail." |
| Final outcome | "Reversals exist but are not linked to the original entry; matching would be by amount and date." | "Outcomes are tracked." |
| Inputs | "Supporting reports are stored as final versions only." | Leaving it out. |
| Task | "Close checklists exist from 2019; earlier tasks are implied by the entries themselves." | "Every task is documented." |
A plain description helps everyone. SourceX can judge whether the remaining evidence supports a verifiable task, a rubric-based review or neither, and the owner is not surprised later. Related reading: failure data for AI training covers why corrected errors matter, and ethically sourced AI training data covers consent and rights.
When a gap ends the conversation and when it does not
Gaps in evidence are a matter of degree. Gaps in eligibility are not. Stop if the company has fewer than 50 full-time employees at peak (contractors excluded), lacks a documented operating history or the rights to license, or has no 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.
Otherwise, describe and continue. A smaller archive with every link intact can be a stronger case than a large one with the outcome missing. Interest in verifiable rewards among vendors and model builders does not mean any particular archive qualifies; why AI companies pay for data sets out what is and is not known. What stays unknown until SourceX reviews it: rights, scope, de-identification of personal and counterparty details, and whether any buyer wants the record type. Partners describe gaps from conversation and never request ledgers or exports to prove them. If the owner proceeds, the review packet lists what the company prepares.
Disclose that you are a SourceX referral partner and may receive a reward, paid separately by SourceX and never deducted from the company's payout, and sign the Channel Partner Agreement before submitting.
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 the gaps are described plainly and the owner still wants to explore licensing, register as a partner, confirm the company on who qualifies and read the program terms.