What went wrong

Defects we found in our own work, what we changed, and how we checked it again.

Check one item could say no recall named something when one did

Found: Code review of the new form, three rounds running. Bodum, sterilizer and BRZ 0083 in the label field, then plurals and a letter wrong, all read as no recall. The rules clear what they can't tie to a product, and the page turned that into a statement about every recall. A crafted link could also put words inside the answer.

Changed: It now also reads CPSC's text directly for what was typed. A miss says we didn't find a recall naming this, which is all a search can say. Nothing typed goes in that sentence. 2d38e21, 3acfc16, 300f593.

Checked again: 787 cases typed from 150 recalls' own text, with plurals and a dropped letter: none read as not found. That's typing built from CPSC's words, not owners' typing, so a miss is still possible.

A size beside an exact model stopped it confirming

Found: Code review found 12000-BTU beside a recalled model read as a second code, after a change of ours the same night. The first fix was too wide, and let 7928-OZ pass beside a recall listing 7927-OZ.

Changed: A size says nothing about which unit this is, unless this recall's own models are shaped that way. a1d2cc7, ffff51c.

Checked again: Sizes added beside the model in 1,374 confirmed cases from CPSC's records: none stopped confirming. 7928-OZ is a near-miss again.

A store check read a shared brand as a reason to pull an item

Found: We ran the check on one real store's public listings. A title that only shared a word with a recall read as if it should come off the shelf, and one card said same line when the shared words only said who the product was for.

Changed: Only a confirmed match now says pull it from the shelf. A shared brand says check the shelf, cards that only share a word sit apart, and that card says same brand. The page says it can't tell whether the store sold a recalled product. ac7ad37, e2ce270.

Checked again: The same store on the deployed site: 238 listings, 1 needs a check, 58 share a word only. One store, so it says little about others.

Pressing Buy twice made two sales

Found: PayPal's webhook can take a minute or two, so the demo's Buy button looked like it had done nothing. A second press made a second sandbox sale.

Changed: The buttons stop once pressed and say it can take a minute or two. The server refuses a second sale while one is on its way, including a request sent straight to it. 7244f56.

Checked again: A test on the deployed site presses Buy, finds both buttons disabled, posts straight to the server and is refused, then sees the buttons come back once the sale arrives.

A CPSC outage read as ours, and an error row was stored as a read

Found: From about Nov 7, 2026, 00:05 UTC, CPSC's API answered HTTP 200 with one fake row, its error in the title. We stored that as a partial read, so /api/health answered 503 and those dates counted as read. The poll starts a day back, so nothing was skipped this time.

Changed: That row is now a failed fetch and marks nothing as read. /api/health answers 200, names CPSC and says since when, and 503 stays for this site's own failures. 9d51025, 6fdda6e.

Checked again: During the same outage, on the deployed site: the demo shop's 9 items, Replay's 934 recalls and its recount, the public check and the MCP tool all answered from the recalls held here. CPSC answered again by 01:30 UTC, and the first good read started from Nov 6, before the outage. It failed again from 02:10 UTC that day, and those reads said so. On Nov 10, 03:43 UTC, a read succeeded again. Still open until a full day passes with no failed read.

A title one character off a model could read as a confirmed match

Found: An outside review ran the real matcher on a title with a code one character from CPSC's model, plus two words from CPSC's description. It came back confirmed, and a confirmed match can lead to a notice and a refund. Found in review of the code, not by tuning on the held-out set.

Changed: A title carrying a code that looks like CPSC's model but isn't exactly it can't be confirmed by its look, so it asks the family instead. 4827c76.

Checked again: New tests cover codes one edit away, and codes that look nothing like a model, such as 2-IN-1. The eval, Replay and the demo queue gave the same numbers before and after.

An uploaded sale could miss a recall the public check found

Found: The public check ignored punctuation in a model number, and the shop check didn't. A title with MY-002 could miss CPSC's MY002 unless similar words happened to save it.

Changed: Both sides now drop punctuation before comparing, in the shop check too. fe43ee0.

Checked again: A test through the real SQL finds the recall with a hyphen or dots on either side. The eval, Replay and the demo queue gave the same numbers before and after.

Asking PayPal again about a refund skipped the dispute check

Found: If a refund's first request never reached PayPal, asking again is the real first refund, and a dispute may have opened since. That path never read disputes, though a first send does.

Changed: Asking again now reads PayPal's disputes first, and sends nothing if there is one or the answer can't be read. de71c0a.

Checked again: New tests with a stand-in PayPal: a dispute stops it, and so does an unreadable answer. Not yet run on the sandbox.

Sandbox refunds started staying PENDING, first from $30, then at $18 too

Found: From Nov 6, 2026, 17:37 UTC, PayPal's sandbox answered refunds of $30 and up with PENDING, reason ECHECK, including calls with no app code in the way. Refunds of $1, $5 and $20 completed. We didn't know why then.

Changed: The refund page says a pending refund isn't finished, and doesn't blame a late webhook. The site now runs on a second sandbox business account whose PayPal balance is funded, with its own app and webhook. The demo's shelf item still costs $18. db96fb8, 1558a0b.

Checked again: On the old account, read Nov 6, 22:39 UTC and Nov 7, 13:28 UTC: refunds of $90, $90 of a $500 sale and $18 all stayed pending, reason ECHECK, while its balance report read $55,956.00, and a $24.00 payout came back INSUFFICIENT_FUNDS. That points to an empty real balance. On the funded account, Nov 7, 20:02 UTC: an $18 sale and its full refund came back COMPLETED, and still COMPLETED 65 seconds later. After the switch, on the deployed site, Nov 7, about 23:23 UTC: a demo shop's $18 sale reached the queue in 16 seconds through the new webhook, its $18 refund read COMPLETED, PayPal's refund webhook was verified, and Replay read the same counts as before. Refunds made on the old account stay pending there.