# Negative demonstrations: a worked example | SourceX

Canonical: https://refer.sourcex.si/guides/concept-negative-demonstrations
Last updated: 2026-11-04

Negative demonstrations are recorded examples of actions that fail, kept together with the reason they fail. In an ordinary purchasing workflow, one looks like an integration that retried a timed-out order as a fresh create and sent the supplier a duplicate purchase order, followed by the ticket, the cancellation and the fix. For a referral, what matters is whether the company kept the state before the action, the action, its consequence and the written reason, and who can authorize a licensing discussion.

## What does negative demonstrations mean in an ordinary business workflow?

It means keeping a mistake on file with its explanation. A negative demonstration is the counterpart to a worked example of the right way: instead of "do this," it says "this looked reasonable, and here is what it broke."

Ordinary business systems produce these whenever someone reverses a mistake and writes down why. A canceled duplicate order, a reversed payment, a voided credit memo with a note, a rollback ticket: each one pairs a wrong action with its consequence. What turns that into a demonstration is the reason. Without it, a buyer sees only that a record was canceled, not which choice caused the problem or what the safer choice would have been.

AI developers care because software agents act through the same systems, and they meet the same traps people do. Our explainer on [why AI agents fail](/questions/why-ai-agents-fail) covers the general pattern. AI data vendors also advertise this kind of work: Appen's agentic AI page, for example, lists failure-mode annotation across plan, tool use and output. That shows the category exists. It does not show that any particular company's error archive is wanted, or usable.

## The duplicate purchase order, annotated (Illustrative)

This case is fictional. A 260-person industrial distributor approves a purchase order for packaging film worth $18,400. Its procurement tool sends the order to the ERP through an integration, and the ERP passes it to the supplier.

| Time | What happened | Record it left | Why it matters to a reviewer |
|---|---|---|---|
| Day 1, 14:02 | Integration sends a create request for the approved order | Integration run log with payload reference | The intended state: one order, $18,400 |
| Day 1, 14:02 | ERP creates PO-55120, but the response never reaches the integration | ERP record with creation time; integration log shows a timeout | The outcome is now unknown to the sender |
| Day 1, 14:03 | Integration retries as a fresh create, with no stable external reference | Second run log; ERP creates PO-55121 | **The failing action.** This is the step a negative demonstration isolates |
| Day 1, 14:10 | Both orders go to the supplier | Two outbound order messages | The error has left the building |
| Day 9 | Buyer sees $36,800 on the supplier statement and opens a ticket | Ticket with both PO numbers | The consequence, written down by a person |
| Day 10 | PO-55121 canceled after the supplier agrees; integration changed to check for an existing order first | Cancellation with reason code; change ticket | The reason and the safer alternative |

The annotation that makes row three useful comes from published guidance, not opinion. The HTTP standard (RFC 9110) classes POST as not idempotent and says a client should not automatically retry such a request unless it has some way to know the retry is safe. Stripe's API documentation makes the same point for payments: a server error leaves the outcome unknown, so check the object before retrying under a new key, and use an idempotency key so a retry cannot perform the operation twice. NetSuite's documentation recommends external IDs with its upsert operation specifically to prevent duplicate records. A reviewer can therefore label the retry as wrong for a stated, checkable reason.

## Evidence fields that turn a mistake into a demonstration

An error on its own is just noise. The duplicate order becomes a negative demonstration only when these fields survive and can be joined.

| Field | Value in the example (Illustrative) | Why a reviewer needs it |
|---|---|---|
| Starting state | One approved order for $18,400, not yet in the ERP | Shows what "correct" would have looked like |
| Action taken | Second create request at 14:03 with no stable external reference | Isolates the exact choice that failed |
| Signal available at the time | Timeout on the first request; outcome unknown | Shows what the actor knew when choosing |
| Consequence | Two orders sent; $36,800 committed | Measures the harm in business terms |
| Detection | Buyer's ticket on day 9 citing both PO numbers | Dates the discovery and names who found it |
| Correction | PO-55121 canceled with a reason code | Shows how the state was restored |
| Written reason | Retry should have checked for an existing order first | Turns the record from "what happened" into "why it was wrong" |
| Safer alternative | Integration changed to look up by external ID before creating | Pairs the failure with the right behavior |

The last two rows carry most of the value. Plenty of companies have canceled duplicates. Fewer have the note that explains the cause and the change that prevented a repeat, which is what lets someone label the failing step rather than guess at it.

What this one example cannot show is how often the pattern recurs, whether every case carries a reason, or how long the integration logs were kept. Those are questions about the archive as a whole, and they are answered in SourceX's review with the company, not from a single case.

## Who can open the conversation, and what stays unknown?

The company needs 50 or more full-time employees at peak (contractors excluded), documented operating history, rights to license its records, 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.

The right first contact is usually the CFO or COO, because purchasing, ERP and integration records sit under them, with the CEO approving. Whether a company can license this material at all is covered in [can a company license its data to AI companies](/questions/can-a-company-license-its-data-to-ai-companies), and the evidence that AI developers pay for business data is collected in [AI companies pay for data](/proof/ai-companies-pay-for-data).

Unknown until SourceX reviews it: how many failures pass the acceptance check, how supplier and employee details are de-identified, whether vendor contracts limit use, and whether any buyer wants this failure type. Partners never request exports, screenshots or sample transactions. Related concepts are in the [data types hub](/data-types).

## Go deeper on negative demonstrations

- [Questions to ask a client about negative demonstrations](/questions/does-negative-demonstrations-data-qualify): a short discovery list for a finance or procurement lead.
- [Record gaps that make negative demonstrations hard to assess](/questions/negative-demonstrations-evidence-gaps): the missing links that most often weaken an error archive, and how to spot them early.
- [Negative demonstrations review packet](/resources/negative-demonstrations-review-packet): what an evidence packet for SourceX's review should contain, item by item.

## 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 the procurement or finance lead confirms that reversed orders carry reasons and that integration logs or tickets survive long enough to explain them, [register as a partner](/register), check the company against [who qualifies](/who-qualifies), and read the [terms](/terms) before you introduce the CFO or COO.

## Sources

- https://www.rfc-editor.org/rfc/rfc9110.html
- https://docs.stripe.com/api/idempotent_requests
- https://docs.oracle.com/en/cloud/saas/netsuite/ns-online-help/section_N3532463.html
- https://www.appen.com/agentic-ai
- https://refer.sourcex.si/terms
