Systems and records
Eligible but unverified: labeling a retail return verification dataset
By SourceX Editorial · Reviewed by Noah Loul ·
9 sections
Short answer
A retail return verification dataset should record two verdicts for every return: eligibility (the item, window and proof meet the policy) and authorization (this presenter may receive this remedy on this tender). When the first passes and the second fails, the right answer is usually partial: refund to the original tender, credit at policy value, hold for evidence or escalate. Records that merge the two keys teach a model the wrong lesson.
Checklist
Eligible but unverified
Key takeaways
- Keep eligibility and authorization as two fields with their own evidence, and derive the remedy from both.
- In 41 French goods sellers' archived sale conditions, 40 state a return eligibility rule, yet no sentence ties an identity document to an ordinary return.
- A refund to the card that paid resolves most identity failures on eligible returns without refusing the customer.
- Name changes, replaced cards, gift receipts and delegates are hard negatives that a naive identity match calls fraud.
- Declined returns have no observed outcome, so use adjudication, appeals and a time-based holdout instead of trusting approval accuracy.
What decision sits between an eligible return and a refund?
Two separate questions sit between a returned item and the money going back: is this return eligible, and is this person entitled to this remedy on this tender? Eligibility asks about the item and the transaction: the window, the condition, the proof of purchase, the category exclusions. Authorization asks about the presenter and the payment route: whether the person at the counter, or the sender of the parcel, is the purchaser or someone the purchaser authorized, and where the refund may go.
The hard cases are the ones where the first key turns and the second does not. A receipt is found, the item is inside the window and unworn, but the name on the identification differs from the order, the presenter declines to show ID, or a return authorization service answers deny. A yes-or-no model trained on refund outcomes tends to file these as either clean approvals or fraud. The useful answer is usually partial: send the money back to the card that paid, offer credit at the value the policy sets, ask for one named piece of evidence, or hand the case to a reviewer.
A retail return verification dataset is the set of records that teaches or tests that decision: each return with both verdicts recorded separately, the evidence behind each, the remedy the policy allowed, the remedy actually given, and what happened afterward. It trains or evaluates a system that proposes a remedy for a person to confirm, not one that accuses a customer.
Eligibility is written down, authorization is not: 41 sale conditions counted
Public return policies spell out eligibility in detail and say almost nothing about who may collect a refund. We tested that on a public corpus: the general conditions of sale archived by Open Terms Archive's France collection, read at commit 73ab95a of its versions repository on October 28, 2026. Open Terms Archive defines this document type as covering warranty, delivery and return of goods or services bought through a monetary transaction.
Of 59 archived documents, 41 belong to goods sellers with a current declaration and a return passage. In 40 of them, at least one eligibility condition (a time limit in days, the item's condition, or a proof of purchase) appears within 300 characters of a return, withdrawal or refund term. In 26, same-means-of-payment wording appears near such a term, and 29 mention store credit or a gift card there.
Identity documents appear in 26 of the 41, but almost never where an item comes back. Classing each sentence that names an identity document by fixed keyword rules, then reading it, 18 documents use it to hand over an order at a pickup point, a store or the doorstep, two use it in a disputed delivery or a delivery-or-return anomaly file, two in a payment check and two for privacy-rights requests. No sentence ties an identity document to an ordinary return or refund. Several sellers, Cdiscount and Nike among them, also describe who may collect an order on the customer's behalf, so a written delegation rule exists at handover but not at return.
Policy text can label eligibility, because the rules are written and dated. It cannot label authorization, because the rule a clerk or a fraud screen applied lives in point-of-sale overrides, ID capture fields and authorization responses. Those operational records are the only source for the second key. The count, by rule:
| Measure (41 documents in scope) | Documents | What it shows |
|---|---|---|
| At least one eligibility condition near a return term | 40 | Window, condition or proof rules are written down |
| Time limit in days near a return term | 40 | The window is the most stated rule |
| Same means of payment wording near a return term | 26 | The remedy route is often fixed by the text |
| Abuse or fraud wording near a return term | 17 | Refusal grounds exist but are not identity checks |
| Identity document named for order handover | 18 | Identity is checked when goods leave, not when they return |
| Identity document tied to an ordinary return or refund | 0 | Authorization at return is not in the public text |
Two keys, six labels: the schema that keeps them apart
Keep the two verdicts in separate fields, each with its own evidence, and derive the remedy from both. A single status such as approved, denied or refunded collapses three facts: whether the return qualified, whether the presenter was entitled, and which route the money took. The figure shows how the pair of verdicts maps to a remedy label, and the field list gives each case the evidence to recompute it.
At a glance
Two keys, six labels: the schema that keeps them apart
The six remedy labels are approve standard (the remedy the policy gives a verified purchaser or an authorized holder such as a gift-receipt bearer), refund to original tender (the money returns to the card or account that paid, never to the presenter), credit at policy value, hold for evidence, escalate and decline. Every case gets exactly one. The field list for each return:
- Eligibility verdict: pass, fail or unknown, with the rule that decided it and the policy version in force on the purchase date.
- Eligibility evidence: receipt or order match, days since purchase, condition grade, category flags such as final sale.
- Authorization verdict: verified, unverified, mismatch, service deny or competing claim, kept apart from eligibility.
- Authorization method: order email, loyalty account, gift receipt, ID inspected, ID scanned, or none, recorded as a method and never as the ID itself.
- Authorization service response, when one exists: the response code and the reason category the retailer received, not the vendor's score.
- Original tender and requested remedy: card or account, cash, gift card or none; refund, exchange or credit.
- Remedy given, the override code if a person changed the system's proposal, and the approver's role.
- Channel and jurisdiction: in store, by post or by carrier pickup; the country or state whose rules applied.
- Later events inside a fixed observation window: purchaser dispute, chargeback, loss prevention finding, appeal, or none.
Eight returns and the label each one earns (illustrative casebook)
One rule order labels every case, so each label can be recomputed and argued with. The rule, first match wins: an ineligible return is declined; an eligibility unknown is held for evidence; a service deny or a competing claim is escalated; a verified presenter gets the standard remedy; an unverified presenter on a card purchase gets a refund to that card; otherwise credit at policy value if the policy allows credit without identification, and a hold for evidence if it does not.
The cases are invented to cover each label and the traps between them, and a script assigns every label from the rule, so no label was chosen by hand. Case C7 is the easiest to mislabel: under Article 13 of the EU Consumer Rights Directive, a trader refunds a distance-sale withdrawal through the same means of payment the consumer used unless the consumer expressly agrees otherwise, so the name on the returned parcel matters far less than the card the money goes back to. The eight cases:
| Case | Eligibility, authorization, tender | Label | Why |
|---|---|---|---|
| C1 Online order returned in store, ID name differs from the order | Pass, mismatch, card | Refund to original tender | Money goes back to the payer, not the presenter |
| C2 Gift recipient with a gift receipt | Pass, verified holder, gift receipt | Approve standard | The gift receipt is the authorization |
| C3 No receipt, presenter declines ID | Pass, unverified, none | Hold for evidence | No tender to route to and no credit rule |
| C4 Receipt found, authorization service answers deny | Pass, service deny, card | Escalate | A person reviews before anyone is refused |
| C5 ID verified, purchaser reported the order not received | Pass, competing claim, card | Escalate | Two people claim the same goods |
| C6 Purchaser verified, 20 days past the window | Fail, verified, card | Decline | A passing identity cannot rescue eligibility |
| C7 EU online withdrawal, parcel sent under another name | Pass, unverified, card | Refund to original tender | The refund route is fixed by law |
| C8 Cash purchase, presenter is not the loyalty member | Pass, unverified, cash | Credit at policy value | Policy allows credit without ID |
Hard negatives: identity failures that are not failures
Many identity mismatches in return records have innocent explanations, and a dataset needs enough of them to teach that. These cases fail a naive identity check yet belong under approve standard or refund to original tender. Leaving them out trains a model to refuse ordinary customers; labeling them as fraud is worse. Hard negatives worth requesting explicitly:
At a glance
Hard negatives: identity failures that are not failures
- A name change after marriage or a legal change, where the order name and the ID differ but the card and loyalty account match.
- A replaced card with a new number on the same account, which breaks a card-number match without changing the payer.
- Gift receipts and gift returns, where the presenter is never the purchaser by design.
- Household members and business buyers, where an employee or spouse returns goods bought on a shared account with a documented delegate list.
- Orders collected by a designated person at pickup, the delegation the French sale conditions describe, which can carry over to the return.
- Marketplace orders, where the retailer's customer record holds the marketplace's relay details rather than the buyer's name.
- The reverse trap: a verified purchaser returning an ineligible item, which a model fixated on identity will approve.
How the outcome label becomes known, and which returns never get one
The remedy label records what was decided; the outcome label records whether the decision held up, and it arrives later or not at all. For an approved refund, the evidence is what happens inside a fixed window after the return: a purchaser who says they never returned anything, a chargeback on the original sale, a loss prevention finding, an appeal, or nothing. Fix the window before you count, and treat returns younger than the window as open rather than clean.
Declined and escalated returns are the gap. Nobody observes what would have happened had a declined return been refunded, so outcome data only exists for the paths the retailer allowed. A model that learns only from observed outcomes will look accurate on approvals and say nothing reliable about refusals. Two partial fixes help: adjudication by a reviewer who sees the full record later, with the reviewer's role and date stored; and appeals, where a reversed decline is the closest thing to a measured false refusal.
Split by time, not at random. Return policies, fraud screens and store procedures change, and a random split lets the model see the policy version it is tested on. Hold out the most recent months, record each case's policy version, and report results per channel, because the evidence for in-store and mailed returns differs.
What a retailer's records can show, and where they stop
A retailer's records can show the decision, the evidence the clerk or the system saw, and the overrides; they cannot show the third-party reasoning, the identity data itself, or the counterfactual for refused returns. Point-of-sale return transactions, receipt lookups, tender records, override codes and loyalty links are the core. E-commerce adds the order, the return merchandise authorization, carrier tracking and the refund method.
Several limits come from outside the retailer. The Consumer Financial Protection Bureau lists The Retail Equation, a company that monitors returns for merchants and reports on product returns and suspected exchange fraud, among consumer reporting companies, and notes the company provides a free report to a consumer who requests it. A retailer that uses such a service typically holds the response it received, not the model behind it, and the response may be governed by the vendor's terms and consumer reporting law. California's Civil Code section 1798.90.1 lets a business scan or swipe a DMV-issued license only for listed purposes and bars retaining or using the information for any other purpose, which argues for keeping the verification method and verdict while leaving the ID data out of any licensed set.
Eligibility rules carry their own dates. California's Civil Code section 1723 requires a retail seller whose policy denies a full cash refund, credit or equal exchange within seven days of purchase to post that policy conspicuously, with listed exemptions, so the posted policy version on the purchase date belongs in the record. In the EU, the withdrawal rules in Article 13 of the Consumer Rights Directive fix the refund route and let a trader withhold reimbursement until the goods arrive or the consumer shows proof of sending them, which explains many holds that look like suspicion.
Rights and privacy come before any of this leaves the company. Removing names does not by itself create a right to license return records; the retailer needs authority over the records, a privacy review of every identity and contact field, and a check of vendor terms for third-party responses. These points are general information, not legal advice, and counsel should review the specific records and jurisdictions.
Asking SourceX for return verification cases
SourceX sources datasets on request rather than holding stock, so a request for return verification cases starts from your task, not a catalog. The request form on its data request page asks for the domain, the source systems, approximate volume, format, geography, language, historical timeframe, licensing requirements including permitted use, and de-identification requirements, with room for additional specifications.
Use that room for the parts this article argues matter: both verdicts kept separate, the six remedy labels or your own equivalents, the hard negatives you need represented, the observation window for outcome events, and an acceptance test such as a time-based holdout scored per label. Say which channels and jurisdictions you need, and say plainly that identity numbers and vendor scores should be excluded. A candidate supplier would typically be a multi-store or omnichannel retailer whose point-of-sale and returns systems keep override codes and policy versions; whether such records exist, can be licensed and meet your test is established during review, not assumed.
How the policy count was made, and what it does not cover
The count read every file named General Conditions of Sale in Open Terms Archive's France versions repository at commit 73ab95a (October 28, 2026), keeping only services that still have a declaration for that document type in the France declarations repository at commit e40568a. That rule dropped three folders that no current declaration covers. An editorial list of services that sell transport, banking, streaming or consumer-to-consumer resale then dropped 14 more (the list also names Salto, already removed), and one goods seller whose archived page had no return passage mentioning goods was dropped by rule, leaving 41.
Text was read as UTF-8 with any byte-order mark stripped, NFC-normalized, lowercased, with typographic apostrophes unified and whitespace collapsed. A flag counts a document if a whole match of its pattern lies within 300 characters of a whole-word return term (rétractation, retour or retours, or a word beginning retourn, renvo or rembours). A return passage counts as about goods if it contains produit, marchandise, colis or article, singular or plural, except article before a legal reference. The flag patterns, accented or not and whole-word: a one-to-three-digit number, quatorze or trente, optionally followed by a number in parentheses, then jour or jours; état d'origine, parfait état, non porté, non lavé or non utilisé (each also with feminine or plural endings), étiquette on its own, or emballage d'origine, the last two singular or plural; preuve d'achat, ticket de caisse, facture or factures, bon de retour, numéro de commande or bordereau; même moyen or mode de paiement, or moyen or mode de paiement utilisé or initial; un avoir, d'un avoir, sous forme d'avoir, bon d'achat, or carte cadeau or e-carte cadeau, singular or plural, with a space or a hyphen; abus, or a word beginning abusi or fraud. The condition and proof flags are each set in 35 documents, counting étiquette on its own and these plural forms.
Identity mentions are pièce, justificatif or carte d'identité, passeport, vérification d'identité, or vérifier votre, son or l'identité. Each was classed by the sentence containing it, split at a period, exclamation mark, question mark or semicolon followed by a space, testing in a fixed order: dispute (litige, anomalie, or a word beginning contest), order handover (retrait, relais, click, numéro or confirmation de commande, or a word beginning retir, récupér, livr or réceptionn), payment check (paiement, solvabilité, bancaire, vérification, or a word beginning fraud), privacy request (droit or droits, données personnelles, opposition, RGPD), return or refund (the return terms above), then other. The script prints every classed sentence, with email addresses removed; it and its output, bin/xs0815_cgv_count.py and .out, are kept in the batch's scratch folder.
Limits: the corpus is French online sale conditions, not in-store return desks, US policies or internal procedures. A keyword window can count a rule that applies to something other than the return, and online conditions naturally say more about delivery than about the counter. The count shows where public text speaks and where it is silent; it does not measure how often retailers check identity at returns, or how often those checks are right.
Frequently asked questions
What is a retail return verification dataset?
It is a set of return records that keeps two verdicts apart: whether the return met the policy and whether the person was entitled to the remedy on that tender. Each record also holds the remedy given, the evidence behind both verdicts and any later dispute or appeal, so a model can be trained or tested on the decision rather than on refunds alone.
Is a failed ID check the same as return fraud?
No. Name changes, replaced cards, gift returns and household or business delegates all fail a naive identity match without any fraud. A failed check is one input to the remedy; the fraud question, which needs investigation and a finding, belongs in a separate label.
Why route a refund to the original card instead of declining?
Because the risk in an identity failure is paying the wrong person, and a refund to the card that paid removes most of that risk while honoring an eligible return. It is also the default route for EU distance-sale withdrawals under Article 13 of the Consumer Rights Directive, unless the consumer agrees otherwise.
Can a retailer license ID scan data collected at returns?
Usually the safer design keeps the verification method and verdict and leaves the identity data out. California's Civil Code section 1798.90.1, for example, limits the purposes for which a business may scan a license and bars other uses of that information. Counsel should review the specific records and jurisdictions.
What does the count of French sale conditions show?
Of 41 goods sellers' archived sale conditions, 40 state at least one return eligibility rule and 26 fix the refund route, while identity documents appear mostly at order handover and never in an ordinary return passage. Policy text can label eligibility; authorization has to come from operational records.
Sources
- Open Terms Archive's France versions repository holds the archived general conditions of sale counted in this article (59 files at commit 73ab95a, October 28, 2026). Source
- The France declarations repository lists which services currently declare a General Conditions of Sale document; it set the population rule. Source
- Open Terms Archive defines General Conditions of Sale as covering warranty, delivery and return of goods or services bought through a monetary transaction. Source
- Article 13 of the EU Consumer Rights Directive requires refunds of a withdrawal through the same means of payment unless the consumer expressly agrees otherwise, and lets the trader withhold until the goods arrive or proof of sending is supplied. Source
- The CFPB lists The Retail Equation among consumer reporting companies as monitoring returns for merchants and reporting suspected exchange fraud, with a free report on request. Source
- California Civil Code section 1798.90.1 lets a business scan or swipe a DMV-issued license only for listed purposes and bars retaining or using the information for other purposes. Source
- California Civil Code section 1723 requires a retail seller whose policy denies a full refund, credit or equal exchange within seven days to post that policy conspicuously, with exemptions. Source
- SourceX's data request page says datasets are sourced on request, not held in stock, and lists the request fields used in the closing section. Source
Tell us what data you need
Share scope, volume, format, timing and licensing needs. Datasets are sourced on request.