Which card did that charge use?

A customer says the wrong card was charged, or says a card was charged that they do not hold. Tonight, three proofs, in a console you administer. Name the card on that charge: the charge id, the payment-method type, the brand, and the last four. Show a different charge whose brand or last four is not this one. Show another account's list that does not contain this charge. If any of the three needs the person who wired checkout, you have not shown the card. You have a receipt total and a guess.

This is the method on one charge. It is not the amount they were shown. Show the price they accepted is the figure on the screen and the figure that was taken. A price can be right while the last four is wrong, and a last four can be right while the amount is not the one they saw. It is not the status of a subscription. Which status is that subscription in? is one of the names a subscription stores. A charge can succeed while the subscription is past due, and a brand is not active. It is not a result that landed twice. Stop the retry that landed twice starts after a correct action happened again. A second charge is a second id, and a second id is not a brand. It is not which endpoint an event reached. Which webhook did that event hit? is a delivery. A delivery can name a charge and still not be the card hash you stored.

None of what follows is a rate for this product, and none of it is legal advice. The fields below are one processor's current description of a charge. The percentages below that are a 2026 conference paper about one processor's chargebacks, after the last four digits had been removed. A brand enum is not a census of your customers. A paper that deleted the last four cannot name the card on the charge you opened tonight. They do tell you which numbers people mix up when they try to sound precise about a card they have not opened.

Stripe's charge object says the status of the payment is either succeeded, pending, or failed. Three names. succeeded is not a brand, and failed is not a failed CVC check. The card checks, on the same object, are pass, fail, unavailable, or unchecked, and they apply separately to address line one, postal code, and CVC. A check of fail is not a charge of failed. A check of pass is not a charge of succeeded. The sample object prints status succeeded beside paid true, captured true, and an outcome type of authorized. authorized is not succeeded. paid is not a status. The sample also prints livemode false, so the object is an example, not a live payment on your account. Its created value is the integer 1679090539. That integer is not the expiration year printed beside it (Stripe, The Charge object).

The card, when the type is card, is a snapshot on payment_method_details, not the id in payment_method. The sample's payment_method is card_1MmlLrLkdIwHu7ixIJwEWSNR. That string is an identifier. The sample's brand is visa, and its network is also visa. Two fields can print the same word and still be two fields. Do not delete network because brand already says visa. The brand list on the card hash is amex, cartes_bancaires, diners, discover, eftpos_au, jcb, link, mastercard, unionpay, visa, or unknown. Eleven names. link is a brand in that list. It is not a URL. unknown as a brand is not the funding value unknown. Funding is credit, debit, prepaid, or unknown. Four names. credit is not visa. The sample's funding is credit, its country is US, its exp_month is 3, and its exp_year is 2024. 3 is a month on an example. It is not a count of proofs. 2024 is an expiration year on an example. It is not a calendar year of losses.

The sample's last4 is 4242. 4242 is four digits on an example object. It is not your customer's card, and it is not a port, and it is not a count of charges. The fingerprint on that same card is mToisGZ01V71BCos. A fingerprint is not a last four. The page says that as of May 1, 2021, the card fingerprint in India for Connect changed to allow two fingerprints for the same card, one for India and one for the rest of the world. Two fingerprints is not two charges. May 1, 2021 is not the sample's exp_year of 2024. The outcome on the sample prints risk_score 32, risk_level normal, and network_status approved_by_network. 32 is a score on an example. It is not a count of merchants, and it is not 4242. A seller message of Payment complete is not the status succeeded, even though both sit on the same sample.

A last four is not always a card. The same object documents ach_debit.last4 as the last four digits of a bank account number, and bancontact.last4 as the last four characters of an IBAN. The field name last4 is shared. The object is not. A bank's last four is not visa. An IBAN's last four is not 4242 unless that example happens to say so, and this example's 4242 is on the card hash, not on the bank hash. The statement descriptor note says a non-card charge can put text on the customer's statement, and it points at a 22-character limit. 22 is a maximum length for a string. It is not 22 percent of anything. The sample's calculated_statement_descriptor is Stripe, which is shorter than 22, and its statement_descriptor is null. A null is not the word Stripe. amount_captured can be less than amount when the capture was partial. amount_refunded can be less than amount when the refund was partial. A smaller amount is not a different card.

The migration page is a second list, and it does not match the object page name for name. It says the older source.brand could be American Express, Diners Club, Discover, JCB, MasterCard, UnionPay, or Visa. Seven display names. American Express is not amex. MasterCard is not mastercard. The same page says payment_method_details.card.brand can be amex, diners, discover, jcb, mastercard, unionpay, or visa. That list does not print link, cartes_bancaires, eftpos_au, or unknown. The object page does. Do not paste the migration list over the object list, and do not paste the object list over a charge you have not opened. The migration page also moves last4 from source.last4 to payment_method_details.card.last4, and it moves a wallet: source.tokenization_method of android_pay becomes card.wallet.type of google_pay. android_pay is not google_pay. Creating charges through the Charges API is deprecated on that page. Reading a charge that already exists is not creating one. A deprecation of create is not a blank brand (Stripe, Charges versus Payment Intents).

The study people reach for when they want a fraud rate is a different document, and it says it threw the card number away. Brayden Killeen, Minh Trung Tran, and Fitrio Pakana, at RMIT, presented it at AISC '26 in Melbourne on February 12–13, 2026. The host data are Novatti Group Ltd. The introduction says Australian card fraud reached a record $913 million in 2024, with card-not-present transactions accounting for 89% of losses. $913 million is a national figure they attribute to a citation. It is not the amount on your charge. 89% is not 0.31%. The dataset is 126,184 merchant transactions from January 2023 through June 2025. January 2023 is not May 1, 2021, and June 2025 is not the conference date in February 2026. The abstract's fraud rate is 0.31%. Section 3.1 prints 394 transactions (0.31%) confirmed fraud and 125,790 (99.69%) legitimate, and calls the imbalance 319:1. 394 plus 125,790 is 126,184. 394 is not 125,790. They also say the last four digits of the card number, masked PANs, and BINs were anonymised before analysis. A paper that removed the last four cannot be the last four on this charge (Killeen, Tran, and Pakana, Chargeback Fraud Detection on Anonymised Merchant Data).

The merchant sentences do not use one percentage for one set. Seven of 32 merchants, which they call 22%, account for 80% of fraud cases. 7 is not 32. 22% is the label they put on 7 of 32. 80% is a share of fraud cases, not a share of merchants. The figure caption says outliers exceed 10%. The prose says two outliers exceed 90%. 10% is not 90%. Most merchants, 25 of 32, have fraud rates below 1%, and the figure text also says most are below 0.5%. 25 plus 7 is 32. 1% is not 0.5%, and neither is the base rate 0.31%. Domestic fraud is described as approximately 0.95% and international as approximately 0.3%, more than 3 times. 0.95 is not 0.3, and 0.3% is not 0.31%. Credit is 0.61% and debit is 0.07%, which they call 8.7 times. 0.61 is not 0.07. 0.07% is not the false-positive rate they later print as 0.06%. Transactions above $5,000 are said to show a 12 times higher fraud rate than transactions under $100. The page does not print those two rates, only the 12. Do not invent the missing pair. A methods flag for high-value transactions, above $1,000, is not the bucket above $5,000.

The model table and the error paragraph do not share one false-positive count. Table 1's neural network prints F1 0.607, precision 0.860, recall 0.468, and 28.2 seconds. The prose says that model identifies 37 of 79 fraud cases, calls that 47% recall, and says it generates 6 false positives from 25,158 legitimate transactions, 0.024%, with precision of 86%. 0.468 is what the table prints. 47% is what the prose prints beside 37 of 79. 0.860 is the cell they also call 86%. 6 is not 15. Section 4.4 says 43 false negatives and 15 false positives, 15 of 25,158, 0.06%. 43 is what that paragraph prints. 79 minus 37 is 42, which the page does not print, so do not replace 43 with a subtraction you prefer. 0.06% is not 0.024%. False negatives average $1,298 and sit across 14 merchants. Detected fraud averages $915 and the true positives concentrate in 7 merchants. False positives average $610 and sit in 4 merchants. $1,298 is not $915, and neither is $610, and none of them is $913 million. 14 is not 7, and 7 is not 4. 28.2 seconds is not 28.2 percent. CatBoost's time on the same table is 32.6 seconds. 32.6 is not 32 merchants, and it is not the sample's risk_score of 32.

Table 2 is a second forest, not Table 1's forest. Table 1's random forest is F1 0.586, precision 0.841, recall 0.449. Table 2's baseline random forest is F1 0.554, precision 0.706, recall 0.456, AUC 0.921, and 15 false positives. 0.586 is not 0.554. 0.841 is not 0.706. SMOTE plus a random forest prints precision 0.427, F1 0.452, recall 0.481, AUC 0.915, and 51 false positives. The abstract says SMOTE degrades precision by a factor of 3.4. 51 divided by 15 is 3.4. 0.706 divided by 0.427 is not 3.4. The factor matches the false-positive column. It does not match the precision column. Do not move the verb onto the column that does not divide that way. Isolation Forest prints F1 0.013, precision 0.013, recall 0.013, AUC 0.808, and 74 false positives, and the prose says it detects 1 of 79. 1 is not 37. 74 is not 51, and 51 is not 15, and 15 is not 6. The prose also gives the neural net an ROC-AUC of 87% and a PR-AUC of 50%. 87% is not Table 2's 0.921. 50% is not 0.31%. Feature importance prints MF-1 at 29.4%, payment method at 11.6%, MF-2 at 11.5%, and card category at 10.6%. 11.6% is a column on a model. It is not a brand. They kept 21 of 41 features, 10 categorical and 11 numeric. 10 plus 11 is 21. 21 is not the 22-character descriptor limit, and it is not 22% of merchants. The smoothing factor in the risk formula is 100. 100 is not the training row count of 100,947. The test split is 25,237, of which 25,158 are legitimate and 79 are fraud. 79 is not 394.

Do the three proofs on one real charge, in the console that actually holds it. First, the binding. Copy the charge id, the type, the brand, and the last four. If the type is not card, write the type and do not borrow visa from a different charge. If the last four is absent, write that it was not stored. Do not infer 4242 because a sample object prints it. Second, the other card. A different charge, with a different last four or a different brand, is not this one. A chart of fraud is not that second charge. If you cannot show a charge that is not this card, you have a counter, not a record. Third, the other account. A list that belongs to someone else does not contain this charge id. If the only list is a single shared table and you cannot say whose charge it is, write that. A tenant you cannot name is the finding. It is not a reason to tell the customer the brand was visa.

Three moves feel like the proof and are a second problem. Refunding the charge because the brand was embarrassing, and then describing the missing row as a cleanup. Quoting 0.31% because it looks precise, while the charge you were asked about is one row, and the study that printed 0.31% had already removed the last four. Asking a model which card was used. The model will pick a plausible name from the eleven. It will not open the charge. A price you have not kept, while the customer is arguing about the amount, is show the price they accepted. Opening that screen does not name the brand.

The path is ordinary when an operator can open one order's checkout without opening the next order's tender as the way to prove it. Meditizer is a storefront with a catalog, a cart, and a secure checkout, synced to an admin that tracks orders, and the stack names Stripe. The shape that matters here is one order you can point at, in an account you administer, distinct from the next order. That case is not a card ledger. Nothing in the description is a headcount, a duration, or a result figure. If the only control on the checkout screen changes every order's tender, you do not have this proof. You have a payment form.

There is a version of this afternoon where hiring the path is the wrong purchase. Nothing is charged, because the product does not take money yet. You have no product. You are still choosing who will build, and the document in front of you is a term sheet. The criteria for that choice sit on the page written for a top SaaS development company. That page is criteria. It does not rank anyone, including us. The only fact in the room is a share from a paper. 0.31% is not your charge, and 4242 is not your charge. A customer has already seen another customer's card. Naming a brand does not close a leak. Or you wanted a lawyer to say whether a network rule applies. Hire the lawyer.

When the charge is real and the console cannot name the card, the missing object is specific. An admin action, in an account you administer, that shows one charge id, one type, one brand, and one last four, shows a different charge that is not that card, and shows another account's list without it. A startup software development company that expects this will have put that action next to the rest of the billing screen, before a stranger's complaint is the way you discover the vendor owns the tender. One that needs a call in order for you to see which card is stored is telling you who holds the charge. If the first slice is a fixed capability, an AI integration sprint is the shape where the acceptance sentence can include the three proofs. If you do not yet know which system stores the charge, a two-week AI readiness assessment maps the systems. The findings are yours. They are not a promise that a paper's rate applies, and they are not 4242.

The runner a job used, and the thinking budget a request carried, are different objects: which runner did that job use? and which budget did that thinking use?. Do not record either on this page. A workflow label is not a brand. A token budget is not a last four.

Before you tell them the card was fine, do the three proofs and write the brand. Name the charge, the type, the brand, and the last four, or write that the card was not stored. Show a different charge that is not that card. Show another account's list that does not hold it. Do not refund the row to avoid the brand, and do not ask a model which of the eleven it was. A paper can tell you what happened after one processor removed the last four. The binding is the charge you can defend.