Evidence

$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.

Counted from the CPSC records and each recall’s proof rule; pro-rated refunds and recalls that state no price are left out, not estimated. The method and the largest recalls

Every claim photo is checked against every photo already claimed, at any firm, by its hashes alone. Here is that check across two firms, recorded with no model call. Both firms and their recalls are fictional samples.

1. 1. At Brightline Home Sample recall BH-2026-01

   

   The photo of the cut cord is verified and paid. The claim keeps one row for later checks: the photo’s hash.

   337bfb9f4e7f3e0f

2. 2. What crosses between firms

   Only a row id that opens nothing at the other firm and the 64-bit hash. No photo, claimant, amount or recall reaches Fernway Home.

   { ref, dhash }

3. 3. At Fernway Home Sample recall FW-2026-01

   The same photo is held for a person before any model call, in 103 ms: “This photo matches one sent with a claim on another recall, so a person at the firm will take a look.”

   With Brightline Home’s hash left out of the pool, the same photo goes on to the model step, so the hold comes from the shared hash.

Recorded by scripts/public/record-network-run.ts into content/public/network-run.json, through the verifier a claim uses (src/lib/verify) with the pool the claim route reads from image_hash; tests/network-run.test.ts runs it again. Photo: Dmitry G, Wikimedia Commons, public domain.

Who may write

Firms connected to Remedy, each with a key tied to its firm (scope hash-pool:register), issued when the firm connects. Claims a firm takes on Remedy add their photo's hash as they are checked. A self-serve sandbox key can check but not register, so the pool holds only photos a firm has handled.

Who may read

Anyone, with or without a key: POST /api/v1/hash-pool/check answers whether a hash is within range of one already in the pool or a published recall picture.

What crosses

Only the 64-bit hashes and an opaque ref. No photo, no name, no contact and no firm: the registering firm is kept in Remedy's own audit ledger, never in the pool row or an answer.

Why a pool across firms. On the CPSC’s 2026 docket on recall claims that are not genuine, a recall administrator reported one person filing 767 registrations under one name and one group registering over 2,200 products at about $100 each, and Truepic reported one device filing over 400 claims with photos from identical vantage points, each claim legitimate when reviewed alone (a recall administrator's comment, June 2026; Truepic's comment, May 2026). A photo reused across firms is only visible to a check that sees both firms’ claims, which is what the hash pool is.

A firm can set a recall to ask for a photo taken now. The claim page’s photo button then opens the phone’s rear camera (capture="environment"), the photo library and the sample photo are not offered, and the claim records that the camera was asked for. It is off on the sample recalls, so a visitor on a laptop can try a saved photo. The claim page with it on

What it does. On a phone, the photo is taken at the moment of the claim, so a picture saved from somewhere else is not one tap away. A recall administrator told the CPSC’s 2026 docket on recall claims that it now often requires live capture (a recall administrator's comment, June 2026).

What it does not prove. The browser treats capture as a request: a desktop browser may still show a file picker, and a camera can photograph a screen or a printout. It does not sign the image the way a capture attestation product does. So every photo still goes through the hash pool, the own-photo check’s reading of screens and prints, and the other five checks, with the setting on or off.

1. 1

   A photo of the product made safe, checked against a rule read from the official CPSC remedy text, then paid through PayPal in the same flow.

   The nearest product verifies that the capture is real and the serial fits, and leaves the decision to the firm's agents without paying; the one patent application on this flow had a person's review as its primary path and was abandoned in 2018.

   Truepic Vision for recall and warranty claimsUS 2016/0350756 A1Real Time Results claim portals

2. 2

   A pool of photo hashes shared across firms, so a photo already paid at one firm is held at the next before any model call, with only two 64-bit hashes and a row id crossing.

   The nearest shared-signal network is built on device and capture signals for lenders and insurers; the nearest recall tool checks one claim live at the moment it is made.

   Truepic Risk NetworkA recall administrator's live photo capture

3. 3

   The refund goes back to the original PayPal purchase inside PayPal's 180-day window, else out as a payout, and a payout nobody accepts becomes a re-notice task in the firm's plan.

   PayPal and Venmo are already reimbursement rails at the largest administrator, so the rail is not new; the step from a verified photo to that money in one flow, and the unclaimed payout turned into a task, are.

   Sedgwick brand protectionCCC Estimate-STP and Snapsheet

Each comparison, with its source and date: How it compares. The two-reviewer design and where it comes from: Photo checks.

Checks and payment

Claude Opus 5.5 reads every remedy into a structured proof rule and checks photos against it; plain code makes the decision and PayPal pays.

A proof ruleRefund to the purchaseAPI docsStatusOperationsHow it is built

The owner's screens

Phone first for owners, every rule and decision with an Info button, Plain and Technical views, and light and dark themes.

HomeTry a photoRecalls

Who it helps

447 recalls since 2024 already refund a photo of the product made safe, and more each year. Their portals pay by check in days; Remedy's PayPal payout lands in seconds.

Who runs it firstThe data methodThe sources

Three ideas

Each firm that joins makes the others safer, the refund goes back the way it was paid, and a payout nobody accepts becomes a task in the recall plan.

One photo, two firmsHow it comparesFor AI agents

Sources and labels

This page and its guided tour, every figure linked to its source file, and what is real and what is a sample, labeled.

About

Every check depends on reading the remedy right. Three readers took the same 130 real CPSC remedies, and each field was scored.

Draft: scored against draft labels; final labels are being checked by hand.

Before any AI call, 7 of 7 recorded photos that needed a closer look were stopped by code: 6 asked for a new photo (the recall’s own pictures, turned and cropped copies, another recall’s picture, a blurry photo) and one photo already used on an earlier claim went to a person at the firm. The hash comparison itself takes a median of 2 ms (0.2 to 93.8 ms); reading the photo and answering takes a median of 98 ms (55 to 200 ms), across 7 recorded runs with no model call. Reading the CPSC’s own photos with that match off, the model was right on 16 of 18 readings.

Every recorded run, the model settings compared, and what the model read

Claim limits, in plain rules

• One refund per person per recall. A second claim on the same recall that shares the payee, the PayPal purchase, the PayPal account that paid for it, or the network and browser goes to a person at the firm. A new email address alone does not reset it.
• Paid with no person, up to a limit. A verified claim at or under $50.00 is paid by the auto rule, up to $350.00 a day for each recall. Past that, verified claims wait in Needs review for a person to approve.
• Only hashes are compared. Each identity is kept as a keyed hash, never in the clear, and no firm sees another firm's claimants.

The sample firm’s settings; each firm sets its own. Counted in code by src/lib/firm/limits.ts and checked again when a claim is paid.

Timings from the owner’s claim flow, run end to end against the PayPal sandbox.

• about 11 s

  for one photo check by Claude Opus 5.5 at the claim path’s setting, the median of the recorded live checks: a median of 11.2 s, 6.4 to 20.7 s across 46 recorded live checks. When the product match is unclear, the two reviewers add a median of 10.7 s (8.6 to 13.9 s in 7 runs).

  Counted from content/verify/sweep-results.json, content/verify/sweep-sample.json, content/presets/recorded/*.json

• 16 to 19 s

  from approval to a PayPal payout reaching SUCCESS, in 12 sandbox runs.

  PayPal sandbox, tests/e2e/claim.mjs and tests/e2e/try-payout.mjs

• At once

  for PayPal to mark a sample-shop purchase's refund to its original PayPal capture COMPLETED after the photo checked out (PayPal says a refund takes 3 to 5 business days to reach the buyer); a second claim on that purchase was refused.

  PayPal sandbox, tests/e2e/claim-shop.mjs

• 98 ms

  median for the free picture match to answer one photo, with no model call: server time inside the check, one claim at a time. The hash comparison itself takes a median of 2 ms (0.2 to 93.8 ms); reading the photo and answering takes a median of 98 ms (55 to 200 ms), across 7 recorded runs with no model call.

  On the laptop that recorded the runs, content/public/verify-runs.json. With 100 claims at once, see Under load.

• 0.95

  right-product score for the CPSC’s own 26125 photo, cut cord not confirmed, with the official-picture match off. Claude also saw it is the CPSC's picture; with the match on, that is answered in code before any model call.

  PayPal sandbox, one live call in 12.3 s, content/presets/recorded/cpsc-example.technical.json

Sandbox payments measured November 5, 2026. Try each one: the presets, the sample shop (buy, claim, and get refunded to the original capture).

• 630

  unit tests pass with no keys and no database, as CI runs them: the proof rules, the six checks and the decision, perceptual hashing, payments and webhooks, the firm console, the API and MCP.

  npm test, November 10, 2026. 61 more need what CI does not have: 21 a Postgres database, 38 a Claude key and 2 PayPal sandbox keys.

• 128

  of 128 page views pass axe (WCAG 2.2 AA, serious and critical) with no sideways scroll: 32 pages, each at phone and desktop width, light and dark.

  npm run e2e:a11y, November 9, 2026

• 8

  tools on the MCP server, each answered by the same code as the API: ping, search_recalls, scan_catalog, check_label, get_proof_rule, verify_photo, open_claim, claim_status.

  src/lib/mcp/tools.ts, checked by node tests/e2e/public.mjs

Every figure in this list is read from content/public/facts.json, written by scripts/facts.ts after the suites run. A unit test fails when the README or TESTING.md print a different one.

Every figure comes from the CPSC Recalls API, pulled Nov 5, 2026: 1,239 recalls from Jan 4, 2024 to Nov 5, 2026. A recall counts as a photo refund when its remedy text mentions a photo, picture or image and a refund (525). The headline then reads each one’s proof rule: 447 ask for a photo of the product made safe and pay a refund in money for it.

Run on the earlier snapshot behind our research note (1,227 recalls to October 29), the same classifier gives the same 523 photo-refund recalls.

The method, with every definition

• 50% vs 6%

  Recalls where the firm contacts its buyers directly see about 50% of products corrected by consumers, against 6% when the recall reaches them through a press release, by CPSC staff's recall data. The CPSC's own workshop report finds direct notice recalls the most effective.

  In Remedy. A firm finds the buyers who paid it with PayPal through Transaction Search, or in its own sales file, and Remedy makes each one a personal claim link and a notice. The firm downloads them for its own mail tool and sends them itself.

  CPSC recall data (2017 workshop), as reported by the OECD, Enhancing Product Recall Effectiveness Globally, 2018

• 72% vs 31%

  In a behavioural experiment for the European Commission, 72% of people shown a direct recall notice decided to act on it, against 31% shown a generic one. Of those who took no action, 21% said the recall process was too much effort or time.

  In Remedy. The claim is one page on a phone, no account and one photo, with the refund named in its first line.

  LE Europe, VVA, Ipsos and ConPolicy for the European Commission, Behavioural Study on Strategies to Improve the Effectiveness of Product Recalls, 2021

• About 70%

  Vehicle recalls, where the maker mails each registered owner, fix about 70% of vehicles within 18 months, by NHTSA's figures. Consumer products have no owner registry, so the purchase record is the closest thing to one.

  In Remedy. PayPal's own record of the purchase stands in for the registry: the buyer is found, and the refund goes back to that same payment inside 180 days.

  U.S. Government Accountability Office, GAO-11-603, Auto Safety, June 2011

• 11.4%

  Across 338 CPSC recalls from 2001 to 2013, the share of recalled products corrected was 11.4% higher on average when the firm offered a full remedy (a free repair, exchange or refund) instead of a partial one, all else equal. Recall effectiveness here is the share of affected products corrected, from the firms' CPSC progress reports. Experiments in the same study find owners see more benefit and feel more able to take part.

  In Remedy. Each claim names the refund in its first line, and the remedy becomes up to four plain steps with a drawing for each.

  Raithel, Hock and Mafael, Journal of the Academy of Marketing Science, 2024

• 14% to 31%

  From a tax credit, not a recall: in an IRS field experiment with 35,050 tax filers who had not claimed the Earned Income Tax Credit after a first notice, a simpler second notice raised claiming from 14% to 23%, and showing the amount they could get more prominently raised it to 31%.

  In Remedy. One page on a phone, no account, and one photo. The checks explain each result in plain words.

  Bhargava and Manoli, American Economic Review, 2015

• 6 in 100

  Recalled products that come back through a recall, on average, by the CPSC's measure: every finished claim adds to it.

  In Remedy. A finished claim takes one photo: the check runs in about 11 seconds (the median of the recorded live checks) and the PayPal payout in about 19, measured in the sandbox.

  CPSC Recall Effectiveness Workshop, 2017, as cited by Raithel, Hock and Mafael, 2024

• 400,000+

  Products covered by the CPSC's January 2025 order to Amazon: email each buyer directly, and refund the full price to anyone who submits proof the product was destroyed or disposed of.

  In Remedy. The same pattern for any firm: buyers who paid with PayPal are found through Transaction Search, and each proof photo is checked and refunded.

  U.S. Consumer Product Safety Commission, January 17, 2025

• 10 bits

  The difference hash (dHash) treats two pictures more than 10 bits apart as different images, and 1 to 10 bits as variations of one, after tests on over 150,000 images.

  In Remedy. A claim photo is matched within 10 bits to its own recall's pictures, and within 4 bits to every other recall's, in a catalog of 1,779 published pictures, across all recalls, distinct enough to tell their recall apart.

  Neal Krawetz, The Hacker Factor Blog, 2013

• 440 million

  Active PayPal accounts at the end of September 2026.

  In Remedy. Refunds go back to the original PayPal purchase inside 180 days, or out as a PayPal payout to an email or phone number, with no bank details asked.

  PayPal Holdings, third-quarter 2026 results, October 27, 2026

• Every photo check, recordedEach recorded run with its six checks, the model settings compared, and what the model read on the CPSC's own photos.
• Cost, load and running it for realHow each check stays cheap, one hundred visitors at once, and what a production deployment changes.
• What is at stake, and who runs it firstThe money method and its largest recalls, how Remedy compares with claim portals, and the 30-day pilot.