What is at stake, and who runs it first

$471.1 million

in refunds these recalls offer at the lowest stated price: 16,705,640 units across the 442 of 447 recalls that refund a photo of the product made safe and whose CPSC record states both units and a price. The two largest recalls, 24309 and 25394, are $184.0 million of the $471.1 million (39.1%), and the median recall offers $68,000.

$261.2 million to $442.8 million

of those refunds stays unclaimed, at two published rates that measure different things: $261.2 million at the 44.55% the CPSC reports as its FY2025 average correction rate across recalls, measured by a method revised in FY2024; $442.8 million if about 6 in 100 recalled units come back, the CPSC’s figure for how many products return after a recall.

$4.7 million

goes back to owners for each extra point of return, one more product in every 100.

The method: for each of the 447 recalls whose proof rule asks for a photo of the product made safe and pays a refund in money, its US units times the refund for one unit. That is the lowest sale price its CPSC record states in the Sold At field, or the fixed refund its rule states when that is lower (4 recalls). 442 records state units and a price. 3 pro-rated refunds (25355, 25323, 25320) state no share of the price and 2 records state no price, so those 5 are left out, not estimated. At the highest stated prices the same refunds come to $1.2 billion. Both rates are averages across consumer product recalls, not measures of these recalls in particular, so the unclaimed figure is shown at each.

The unclaimed range is an inference: it applies rates measured across recalls to these recalls' dollars, not a count of unpaid refunds.

A stated refund above the lowest price counts at that price, so the figure stays a floor: 24309 states a $40 refund, more than its lowest stated price, and counts at $35.00. Written by scripts/recalls/value.ts to content/recalls/value.json from the CPSC Recalls API pull and the proof rules; tests/recall-value.test.ts recomputes it.

What the regulator, the firms and the payment issuers report, from their own pages and filings, and what Remedy does about each.

• 44.55%

  The CPSC's own performance measure for FY2025: an average correction rate of 44.55% across all consumer product recalls, against its 43% target. The method was revised at the FY2024 midyear review, so this figure is an average across recalls; the 6 in 100 above is a different measure, the share of recalled units that come back, and the two are not comparable.

  In Remedy. Every finished claim is a correction with a timestamp, so the firm's monthly progress report to the CPSC is counted from the ledger, not assembled by hand. src/lib/firm/report.ts, src/lib/audit.ts

  CPSC, Agency Performance Report, Fiscal Year 2025, April 2026

• 91 FR 20138

  On April 15, 2026 the CPSC opened a request for information on recall claims that are not genuine (docket CPSC-2026-0067). It names coordinated networks working through online submission forms, artificially generated photos offered as proof of destruction, and the risk that heavier verification lowers participation by the owners the recall is for.

  In Remedy. Both halves of that notice: the checks that catch a copied or reused picture run in code before any model call, and the claim stays one photo on a phone with no account. src/lib/verify/dhash.ts, src/lib/verify/quality.ts, src/lib/verify/decide.ts

  Federal Register, Request for Information Regarding Consumer Product Recall Fraud, docket CPSC-2026-0067, April 15, 2026

• 767 and 400+

  Two comments on that docket. A recall administrator reported one group registering over 2,200 recalled products at about $100 each and one person filing 767 registrations under one name, said that many unique photos of destroyed units can be made in seconds, and now often requires live photo capture. Truepic reported one device filing over 400 claims, photos taken from identical vantage points across hundreds of claims, and each claim looking legitimate when reviewed on its own.

  In Remedy. The shared hash pool: every claim photo is compared by its 64-bit hashes with every photo already claimed at any firm and with every published recall picture, so a reused photo is held before a person sees it, and only the hash crosses between firms. src/lib/verify/dhash.ts (CATALOG_MAX_DISTANCE), content/public/network-run.json

  Realtime Results LLC (Recall Results), public comment CPSC-2026-0067-0013; Truepic Inc., public comment CPSC-2026-0067-0010, June 8, 2026; May 29, 2026

• PayPal or a gift card

  Firms already pay photo-proof refunds by PayPal. A fan recall page from June 2026 asks for two photos and no receipt, offers PayPal or a gift card, says PayPal is fastest, promises an email confirmation within 48 to 72 hours, and states no refund delivery time. A 2025 photo-refund page asks owners to allow 4 to 6 weeks; a claim portal run by a recall administrator pays by check in 10 to 15 business days.

  In Remedy. Remedy's PayPal payout reached the owner in about 19 seconds after approval, measured in the PayPal sandbox, and PayPal marked the refund to the original PayPal purchase COMPLETED at once. src/lib/payout-timing.ts, content/public/measured.json, content/public/e2e-transcript.json, src/lib/payments/router.ts

  Merkury Innovations, Hot + Cool fan recall page; Deale International, CasaClean steam cleaner recall page; Real Time Results, Cuisinart grill recall portal (recallrtr.com), its published page text, recall dated June 11, 2026; page read November 8, 2026; July 17, 2025

• 1 of 10,110

  Across all 10,110 records in the CPSC's recall listing since 1973, one record names PayPal (recall 24-002) and none names Venmo; 4 name an electronic payment. The rail firms use on their own pages is nearly absent from the recall notices themselves.

  In Remedy. The count is reproducible: the script reads the CPSC's CSV and writes only the counts, with the file's URL and the date, into the repo. scripts/recalls/listing-mentions.ts, content/recalls/listing-mentions.json

  CPSC recall listing CSV (every recall since 1973), counted by scripts/recalls/listing-mentions.ts, read November 8, 2026, latest recall October 22, 2026

• $0.25 vs $2 to $4

  The issuers' own terms. A PayPal payout through the Payouts API costs the sender a flat $0.25 in the US and the recipient nothing; an unclaimed payout returns to the sender after 30 days; a refund inside 180 days of the purchase reaches the buyer in 3 to 5 business days. A paper check costs its issuer $2 to $4 to send.

  In Remedy. The router follows the 180-day rule, and the 30-day expiry is what the PAYMENT.PAYOUTS-ITEM.UNCLAIMED webhook turns into a re-notice task in the firm's plan, so the owner is reached again before the money comes back. src/lib/payments/router.ts (planPayment), src/lib/payments/events.ts (reNoticeTask)

  PayPal, Merchant fees (US); PayPal Developer, Payouts frequently asked questions; PayPal Developer, Refund a payment; Association for Financial Professionals, 2022 Payments Cost Benchmarking Survey (press release), last updated November 4, 2026; last updated October 14, 2026; last updated June 4, 2026; AFP January 25, 2022

• 98% vs 77%

  In the FTC's study of 149 consumer class settlements, the median claims rate was 9%, and a weighted 77% of the checks sent were cashed (median 94%). A payments processor's 2024 report across nearly 600 cases puts digital payment success at 98%, against 77% of checks cashed where there was a claims process and 55% where there was not.

  In Remedy. The refund lands in the PayPal account the owner already uses, or goes back to the card or balance that paid, so there is no check to cash. src/lib/payments/router.ts (chooseRoute)

  Federal Trade Commission, Consumers and Class Actions: A Retrospective and Analysis of Settlement Campaigns; Digital Disbursements (Western Alliance Bank), 2024 Annual Report: Digital Payments in Class Actions and Mass Torts, September 2019; April 2024

• 70% and 35.1%

  Owners' side. In Consumer Reports' survey of 1,010 US adults, almost 70% had not heard of a recall for anything they own in five years, and 21% had heard of one and responded. In the European Commission's survey of 24,039 people, 35.1% of owners who saw a recall for their product did not react, and 34.3% of those who contacted the firm were offered a reimbursement. The CPSC's 2003 review reports that 23% of non-responders in a vehicle recall survey gave reasons of effort or motivation, such as no time or inconvenience, and about 30% asked for an easier process.

  In Remedy. The notice and the claim are one link: the firm's buyers are found through Transaction Search, each gets a personal claim link with the refund named in its first line, and the claim is one photo with no account. src/lib/payments/buyers.ts, src/lib/firm/notices.ts

  Consumer Reports, When Recalls Fail (nationally representative survey of 1,010 adults); Ipsos for the European Commission, Survey on consumers' behaviour and perception of product recalls; XL Associates for the CPSC, Recall Effectiveness Research: A Review and Summary of the Literature on Consumer Motivation and Behavior, October 31, 2019; April 2019; July 2003

The two reviewers both run on Claude Opus 5.5 today. Their independence comes from the rulebook each reads: one holds the photo against the recall’s official pictures, the other against the remedy text, and when they disagree a person at the firm decides. Each reviewer’s model is its own setting, REMEDY_REVIEWER_MODEL_PRODUCT and REMEDY_REVIEWER_MODEL_REMEDY (read by reviewerModel in src/lib/verify/reviewers.ts, empty for Opus 5.5), so running one reviewer on a second model is a configuration change with no code change; tests/reviewer-model.test.ts checks that each request carries the model its setting names. That change has been measured once, with the Remedy reviewer on Claude Sonnet 5.5: it paid nothing wrongly, and it held back the genuine photo every time, so both stay on Opus 5.5 (the comparison).

What each photo is for

525 recalls since 2024 name a photo and a refund in their remedy. Their proof rules say what the photo must show, and that is what the headline counts:

• 452 ask for a photo of the product made safe: 447 pay a refund for it, the headline, and 5 pay store credit or a gift card (26434, 26053, 25135, 25042, 24341).
• 71 ask for a photo for another reason: 68 of the product itself, as proof of ownership, and 3 of a label, lot or date code (26741, 26702, 24189).
• 2 need no photo once read closely (27015, 24193): the photo is one proof of purchase among others, or the owner visits a store.

The 71 still fit Remedy’s checks. Right product compares the photo with the recall’s own pictures, Your own photo keeps out a copied or already-claimed picture, and Made safe passes because the rule asks for no change (evaluateC2 in src/lib/verify/checks.ts). For the 3 that ask for a label, lot or date code, Within the recall reads the code from the photo and compares it with the range the recall states (evaluateC4 and src/lib/verify/identifiers.ts). They are counted apart because a photo of the product as it is proves ownership, while a photo of it made safe also proves it is out of use.

Written by scripts/recalls/value.ts to content/recalls/proof-split.json from the recalls and their proof rules (src/lib/recall/made-safe.ts); tests/recall-value.test.ts recomputes it.

Each of these does its job. Remedy adds the instant check and the PayPal payout to what firms already do.

1. [1] Real Time Results, Cuisinart grill recall portal (recallrtr.com), its published page text
2. [2] Real Time Results, Bath Shark recall portal (recallrtr.com), its published page text
3. [3] Sedgwick brand protection, automotive recall services
4. [4] Sedgwick brand protection, product recall solutions
5. [5] Merkury Innovations, Hot + Cool fan recall page (recall dated June 11, 2026; page read November 8, 2026)
6. [6] Deale International, CasaClean steam cleaner recall page (July 17, 2025)

Three groups already handle recall claims. Each can run Remedy as it is today.

• Small firms

  80.9% of recalls cover fewer than 50,000 units

  1,002 of 1,239 recalls since 2024, and 465 of the 525 photo-refund recalls. A firm works its claims in the console and pays them through PayPal; the monthly report is drafted for it. Each firm can pay from its own PayPal account.

  The firm console

• Marketplaces

  80% of recalls were sold online

  991 recalls list an online seller. A marketplace can show each recall's checklist on the order page, verify the buyer's photo with one call, and refund the original PayPal purchase inside 180 days.

  The API

• Recall service providers

  46 recalls name a professional recall vendor

  A provider handling many firms' recalls can send each emailed photo to the verify endpoint, keep its own process for the rest, and pay the refunds through PayPal.

  The verify endpoint

That a faster refund raises claims is an inference from the cited studies (a full remedy against a partial one; a simpler tax-credit notice), not a measurement of Remedy.