# Synthetic-to-real validation workflow example | SourceX

Canonical: https://refer.sourcex.si/guides/synthetic-to-real-validation-workflow-example
Last updated: 2026-11-04

In an ordinary workflow, synthetic-to-real validation means taking cases a generator invented, such as simulated return requests, and checking them against two things only the business holds: the records its real cases left behind at each step, and the judgment of the person who handles those cases. The check works only where steps leave decision records: who decided, under which rule, with what outcome. This page follows one illustrative returns authorization workflow end to end and marks which of its records carry that signal.

It extends the main answer to [whether synthetic data will replace real data](/questions/will-synthetic-data-replace-real-data), which compares a simulated return mix with a real one. Here the unit is the workflow itself: every step, every record.

## The workflow, start to finish

Illustrative and fictional. A 310-person industrial parts distributor sells bearings, belts and gearboxes to maintenance teams. Its returns authorization process runs across a customer portal, its ERP, a warehouse system and a supplier portal. Each step makes one decision and leaves at least one record.

| Step | Decision made | Record left behind | Who decides |
|---|---|---|---|
| 1. Request | Is this a return, a warranty claim or a misship? | Request with reason code, order line and serial number | Customer, checked by a returns coordinator |
| 2. Eligibility | Approve, deny or route to warranty, under which policy version | Approval note citing the policy and date | Returns coordinator |
| 3. Authorization | Which items, quantities and return method | Authorization number and expected-items list | System, from step 2 |
| 4. Receiving inspection | What condition did the item arrive in? | Condition grade, inspector notes, photo reference | Warehouse inspector |
| 5. Disposition | Restock, refurbish, scrap or return to supplier | Disposition code with the decider's user ID | Returns lead |
| 6. Credit or replacement | Full credit, partial credit, replacement or none | Credit memo or replacement order in the ERP | Finance, from step 5 |
| 7. Supplier claim | Can the cost be recovered from the manufacturer? | Vendor claim with fault code and recovery outcome | Purchasing |

Seven steps, at least seven records. Only some of them say how a person exercised judgment.

## Running one simulated case through the same steps

Now picture an AI developer generating simulated returns cases from the distributor's published returns policy, and asking the returns lead to walk one through.

Simulated case: a gearbox returned 70 days after delivery, described as "noisy under load", packaging opened.

| Step | What the generator decided | What the real workflow does | Record that settles it |
|---|---|---|---|
| Request | Coded as "customer changed mind" | Noise under load is coded as a suspected defect | Reason codes on past gearbox requests |
| Eligibility | Denied: outside a 60-day return window | Suspected defects route to warranty regardless of the window | Approval notes citing the warranty policy |
| Receiving inspection | Skipped | Every gearbox is bench-tested before disposition | Inspection grades on returned gearboxes |
| Disposition | Scrap | Units with bearing noise go back to the manufacturer | Disposition codes and the decider's ID |
| Credit | None | Credit issued once the manufacturer accepts the claim | Credit memos linked to vendor claims |
| Supplier claim | Not considered | Claim filed under the fault code for bearing noise | Vendor claim outcomes |

The generator followed the policy text faithfully and still got almost every step wrong, because the policy describes the window, not the routing rules people actually apply. Each correction came from a record the distributor already keeps.

## Which records carry the signal

Not every record is equally useful for validation. The test is simple: does the record show a decision, the rule behind it and the result?

| Record | Signal for validation | Why |
|---|---|---|
| Eligibility approval note with policy version | High | Shows the decision and which rule applied at the time |
| Inspection condition grade | High | Records an expert judgment about the physical item |
| Disposition code with decider | High | Shows the choice among options and who made it |
| Credit memo linked to the case | Medium | Confirms the financial outcome; little on reasoning |
| Vendor claim and recovery outcome | Medium | Shows whether the judgment was upheld by a third party |
| Request with reason code | Medium | Gives the real mix of cases; codes are often misapplied |
| Authorization number and shipping label | Low | Links the case together; no decision |

A workflow whose high-signal records survive, with dates and policy versions, is one where simulated cases can actually be checked. That is the property worth describing to SourceX, as the explainer on [what workflow data is](/proof/what-is-workflow-data) sets out more generally.

## Two checks, both ending at real records

Validation has a statistical half and a judgment half. Researchers describe fidelity as statistical resemblance to real data, and commonly measure usefulness by training on synthetic data and evaluating on real data ([arXiv](https://arxiv.org/abs/2410.03411)). In this workflow the statistical half compares, say, the share of suspected-defect requests in the simulated set with the share in a year of real requests. The judgment half is the step-by-step walk above.

Leaning on generated data without either check carries a documented risk. A 2024 Nature study found that indiscriminately training generative models on generated content degrades their output over successive generations, with the tails of the original distribution disappearing ([Nature](https://www.nature.com/articles/s41586-024-07566-y)). In a returns workflow, the tails are the awkward cases: the out-of-window warranty, the misship that looks like a defect. Those are exactly the cases the generator above mishandled. The limits are covered in more depth on [the limits of synthetic data](/proof/synthetic-data-limits) and in the comparison of [synthetic environments and real business logs](/compare/synthetic-data-vs-real-data-ai-agents); a plain definition is in [what synthetic data is](/questions/what-is-synthetic-data).

## Sizing a client's workflow at metadata level

An introducer never asks for the records. Ask about them, step by step, using the table above as a prompt sheet:

- For each step, which system holds the record and from what year?
- Do eligibility notes cite a policy version, and do old policy versions survive?
- Are inspection grades and disposition codes stored as fields, or only in free text?
- Can a request be linked to its credit memo and any vendor claim?
- Who owns the returns systems, and which officer could authorize a licensing discussion?

Before introducing a company, check the baseline: 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. What stays unknown until SourceX reviews it: whether any records are selected, the rights review and any price. The same step-and-record approach for shipping exceptions appears in the [logistics workflow example pack](/resources/logistics-workflow-example-pack), and other record types are listed under [data types](/data-types).

## 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 a client's workflow keeps its eligibility, inspection and disposition records with dates and policy versions, [register as a partner](/register), check the company against [who qualifies](/who-qualifies) and read the [terms](/terms) before you introduce its owner.

## Sources

- https://arxiv.org/abs/2410.03411
- https://www.nature.com/articles/s41586-024-07566-y
- https://refer.sourcex.si/terms
