Systems and records
Retail seller dispute dataset: two sellers, one fulfillment event
By SourceX Editorial · Reviewed by Noah Loul ·
10 sections
Short answer
When two sellers disagree about the same fulfillment event, split the disagreement into separate facts: whether the package left, when, with which items, and who shipped it. Settle each fact from the record created closest to the physical event, such as a carrier scan or pack record, rather than from either seller's status field. Keep corrections beside originals, record the rule versions in force, and confirm the result against the customer outcome.
License terms
Retail seller dispute
Key takeaways
- Split a fulfillment dispute into separate facts: shipped, when, what, by whom, delivered.
- Rank evidence by how close it was created to the physical event; status fields rank last.
- Keep corrected events beside their originals with the declaration time and reason.
- Record the promised window, inventory state and replacement orders in force on the order date.
- Label disputes without independent evidence as unresolved instead of forcing a verdict.
How should a marketplace task handle seller disagreement about a fulfillment event?
Treat the disagreement as several factual questions, each with its own evidence authority, instead of one verdict about which seller is right. A retailer and its drop-ship supplier, or two sellers whose items left in one consolidated shipment, often agree on most facts and conflict on one: the ship date, the quantity, or whether the package was ever handed to a carrier.
For each fact, the record produced by the event itself outranks a party's later description of it. A carrier acceptance scan says more about when a package left than a shipment confirmation typed the next morning, and a pack station scan says more about contents than an order line. Status fields sit at the bottom, because they summarize claims.
A retail seller dispute dataset built this way teaches a model the method a good operations analyst uses: decompose, rank the evidence, decide per fact, then check the decision against what happened to the customer. A dataset that only stores the final status teaches it to trust whoever updated the record last.
What are the two sellers actually disagreeing about?
Name the disputed fact before collecting evidence, because each fact is settled by different records. Most seller-side fulfillment disputes fall into six patterns, and a single case can contain two of them.
| Disagreement | Typical claim on one side | Typical claim on the other | Evidence that usually decides it |
|---|---|---|---|
| Shipped or not | Supplier confirmed shipment with tracking | Retailer shows the order unshipped | Carrier acceptance or first scan for that tracking number |
| When it shipped | Shipped inside the promised window | Shipped after the window closed | Event time of the first carrier scan, with time zone |
| What shipped | Both units packed | Customer received one unit | Pack station scans and package weight at acceptance |
| Who shipped | Supplier sent it directly | Retailer fulfilled from its own stock | Ship-from location on the label and the inventory decrement |
| Delivered or lost | Carrier shows delivered | Customer reports nothing arrived | Carrier delivery event and its location detail |
| Cancelled before or after shipping | Cancellation came first | Package was already in transit | Timestamps of the cancel request and the first carrier scan |
Why can't the order status settle the dispute?
Because the status is downstream of the sellers' own claims. In Amazon's Selling Partner API orders model, an order is Shipped when all items in it have been shipped and PartiallyShipped when some have, and a seller-fulfilled shipment is confirmed by sending package details: carrier code, tracking number, ship date and the order items with quantities. The status reflects what the seller reported.
The vendor direct fulfillment model shows the same pattern from the supplier side. A shipment confirmation carries a shipped date described as the departure from the vendor's location, which should not be in the future, and a shipment status that is either SHIPPED or FLOOR_DENIAL, the latter for orders rejected because of product quality problems or a mismatch between physical and virtual inventory. Vendors that use their own carriers send status updates whose tracking number must match the one in the confirmation.
None of these fields is false by design, but each is a statement by an interested party. A dispute dataset should store them as claims with an author and a time, then link each to the independent evidence that supports or contradicts it.
Which record has authority for each fact?
Rank evidence by how close it was created to the physical event and by who created it. GS1's EPCIS event model is a useful frame: its event schema records what happened as a business step (shipping, receiving, departing, arriving and others), when it happened as an event time with a required time zone offset, and when it was recorded as a separate record time. Keeping event time and record time apart is what lets a resolver see that a confirmation was entered hours after the package left.
The ladder below is a starting point for retail operations, not a standard. A team should adjust it to its own systems and write the adjusted version into the task brief so every annotator applies the same order.
| Fact | Strongest evidence | Supporting evidence | Weakest evidence |
|---|---|---|---|
| Package left custody | Carrier acceptance or first origin scan | Shipment confirmation with matching tracking number | Order status set to shipped |
| Ship time | Carrier scan event time with offset | Warehouse dock or manifest close time | Ship date typed into a confirmation |
| Contents and quantity | Pack station item scans | Package weight recorded at acceptance | Order line quantity |
| Ship-from party | Label origin and inventory decrement at that site | Purchase order routed to the supplier | Seller's statement in a message |
| Delivery | Carrier delivery event with location detail | Customer confirmation | Estimated delivery date |
| Cancellation timing | System log of the cancel request with time | Customer service note | Final order status |
How should corrected or withdrawn fulfillment events be kept?
Keep the original event, add the correction as its own record, and link the two. EPCIS handles this with an error declaration attached to the original event: it carries a declaration time, an optional reason from the standard vocabulary (did_not_occur or incorrect_data) and the identifiers of corrective events that replace it.
Disputes are often where corrections cluster. A supplier voids a label and reprints it, a retailer reverses an inventory decrement, a carrier adds a missing scan a day late. If the dataset overwrites the original, the model never sees that the first story changed, and the timing of the correction is frequently the most informative fact in the case.
For each corrected record, store who declared the error, when, the reason, and which record replaced it. A resolver that ignores a declared error, or treats a correction issued after the customer complained as equal to an original scan, should be graded as wrong.
Which promised window, inventory state or customer outcome changes the valid resolution?
The same facts can support different resolutions depending on the commitments in force. The ship window is the first: the orders model exposes the earliest and latest ship dates a seller committed to, and the FTC's Mail, Internet, or Telephone Order Merchandise Rule requires sellers to have a reasonable basis for the shipping time they state, or 30 days if they state none, and to obtain the buyer's consent to a delay or refund the unshipped merchandise. Whether a shipment was late depends on which promise applied on the order date.
Inventory state is the second. A floor denial means the supplier could not ship because physical and recorded stock disagreed; a later shipment from the retailer's own stock turns one order into two fulfillment paths. Replacement orders are the third: the orders model flags replacement orders and links them to the order they replace, so a dispute about the original may already have been resolved by the replacement.
The customer outcome is the check. Under Regulation Z, a cardholder can assert a billing error for property not delivered as agreed, and the official commentary says the cardholder need not first try to resolve the problem with the merchant. A seller dispute can therefore end in a card dispute that neither seller started, and the packet should link that outcome when it exists.
The evidence authority and dispute resolution task brief
Each task in the dataset is one packet. The packet holds the starting state, both parties' claims, every linked evidence record, the rules in force, the resolution per fact and the verified outcome. The field groups below can be used as the brief's schema; identifiers are pseudonymized before any packet leaves the company.
At a glance
The evidence authority and dispute resolution task brief
| Field group | Fields | Why it matters | Validation check |
|---|---|---|---|
| Case header | Case id, channel, order id (pseudonymized), order date, fulfillment path | Fixes which promise and path apply | Order date precedes every event |
| Parties | Party id, role (retailer, supplier, consolidating seller, carrier), ship-from site | Separates who claimed from who acted | Every claim and event names a party |
| Claims | Author, time made, asserted fact, asserted value, channel of the claim | Keeps statements apart from evidence | Each claim maps to one disputed fact |
| Evidence records | Type, source system, creator, event time, record time, time zone offset, linked fact | Lets the resolver rank evidence | Event time is not later than record time |
| Corrections | Original record id, declaration time, reason, corrective record ids | Preserves how the story changed | Every correction points to an existing original |
| Rules in force | Promised ship window, delay notice sent, return and cancellation policy version | Decides whether a fact is a breach | Rule version effective on the order date |
| Inventory state | Stock snapshot at allocation, floor denial flag, substitutions | Explains split or failed fulfillment | Snapshot time precedes the ship event |
| Resolution | Decision per fact, authority used, confidence, unresolved flag | The label the model learns | Every disputed fact has a decision or an unresolved flag |
| Outcome verification | Refund, reshipment, replacement order link, card dispute result | Confirms the decision held | Outcome dated after the resolution |
Worked example (illustrative): a drop-ship order with one unit missing
All details are illustrative. A retailer sells 2 units of a product on a marketplace and routes the order to a drop-ship supplier. The supplier's confirmation says both units shipped on day 2. The retailer's system still shows the order unshipped on day 3, past the latest ship date, so its team cancels and refunds. On day 5 the customer reports receiving 1 unit.
Decomposed, the case has three facts. Shipped or not: the carrier's first scan exists, so the package left; the supplier is right on this fact. Ship time: the first scan's event time is on day 3, not day 2, and the confirmation's record time shows it was entered after that scan, so the shipment was late against the window and the supplier's ship date is contradicted. Quantity: the pack station recorded 1 item scan and the acceptance weight fits 1 unit, so the confirmation's quantity of 2 is contradicted.
Resolution per fact: shipped, late, 1 unit. The verified outcome is a refunded order with 1 unit delivered, which the packet records rather than smoothing over. The counterexample matters as much: had the supplier used its own carrier with no independent scan, the ship fact would rest only on self-reported status updates, and the correct label would be unresolved with low confidence, not a guess.
Which dispute records should stay out of a licensed dataset?
Exclude what the company has no right to share and what the task does not need. Buyer names, addresses, phone numbers and message text identify people and add nothing to evidence ranking once orders are pseudonymized and locations are generalized to region or carrier facility.
Rights come next. A retailer's own order and inventory records are different from a supplier's pack records it received under a supply agreement, and both differ from data pulled through a marketplace's developer interface, which comes with that marketplace's own terms. Having access to a record through an integration is not permission to license it; each source needs its own check.
Finally, hold back cases still in active litigation or under a carrier claim, and cases whose outcome is not yet known. A packet without a verified outcome teaches the model a resolution nobody confirmed.
Licensing fulfillment dispute histories with SourceX
Dispute histories with linked carrier, pack and inventory records are rarer than raw order exports, which is why they are worth preparing carefully. A retailer or supplier that holds such histories can license them for AI through SourceX, which runs the rights review, contracting, preparation and delivery.
Checking eligibility starts with a short company assessment and needs no upload. Large order archives can stay in the seller's own storage or ship on encrypted drives, and the company approves the scope and signs before any record is delivered.
Frequently asked questions
What should the evidence authority and dispute resolution task brief include?
A case header, the parties and their roles, each claim with its author and time, every evidence record with event and record times, corrections linked to originals, the rules in force on the order date, a resolution per disputed fact, and the verified customer outcome. Pair it with a written evidence ladder so every annotator ranks evidence the same way.
Is a carrier tracking number proof that an order shipped?
Not by itself. A tracking number in a shipment confirmation shows a label was created and reported; a carrier acceptance or origin scan for that number is what shows the package entered the carrier's custody. Store both, and rank the scan higher.
Which ship window, stock or replacement record can change the resolution?
The promised ship window in force on the order date, any delay notice and the buyer's response, a floor denial or stock mismatch that split fulfillment, and a replacement order linked to the original. Each can change whether a late or partial shipment was a breach, and the packet should record them as dated facts.
Should unresolved disputes be included in a training dataset?
Yes, labeled as unresolved with the reason, such as no independent carrier scan. Leaving them out teaches a model that every dispute has a confident answer; forcing a label teaches it to guess.
Can a retailer license records it received from its drop-ship supplier?
Only if its agreement with the supplier allows that use. Having the records in the retailer's systems is access, not permission; the supply agreement, and any marketplace terms covering data pulled through its interfaces, decide what may be licensed.
Sources
- Amazon's Selling Partner API orders model defines order statuses such as Shipped and PartiallyShipped, the package details a seller sends to confirm a shipment, committed ship dates, and replacement order links. Source
- Amazon's direct fulfillment shipping model defines the shipped date as departure from the vendor's location, the SHIPPED and FLOOR_DENIAL statuses, and status updates whose tracking number must match the confirmation. Source
- GS1's EPCIS 2.0 event schema records event time, record time and a required time zone offset, and attaches error declarations with a declaration time, reason and corrective event identifiers. Source
- The EPCIS error reason vocabulary has two standard values: did_not_occur and incorrect_data. Source
- The EPCIS business step vocabulary includes shipping, receiving, departing, arriving, packing and picking, among others. Source
- The FTC's business guide explains that sellers need a reasonable basis for stated shipping times, or 30 days if none is stated, and must obtain consent to delays or refund. Source
- Regulation Z section 1026.13 treats a charge for property not delivered as agreed as a billing error a cardholder can assert. Source
- The official commentary to section 1026.13 says a consumer need not first try to resolve a non-delivery problem with the merchant before asserting a billing error. Source
See if your company qualifies
A short company assessment. No data uploads are needed.