How well it works
On held-out recalls, when a title can't settle it, Recall Reach asks instead of guessing in 11 of 12 pairs. A plain word match guesses in 11 of 12.
| Held-out pairs | Recall Reach | Word match |
|---|---|---|
| A look-alike reaches the seller | 37 of 83 (45%) 34 to 55% | 45 of 83 (54%) 44 to 65% |
| A look-alike could reach a family | 2 of 83 (2%) 1 to 8% | 45 of 83 (54%) 44 to 65% |
| Claude labelled all 285 pairs, from 75 recalls. One labeller. No person has audited them. A second model from OpenAI (gpt-6.1-sol), blind to Claude's labels, checked 50 of them: it agreed on 34, said the title can't settle it on the other 16, and called none of Claude's look-alikes the recalled product. The files. | ||
| Title can't settle it: asks | 11 of 12 (92%) 65 to 99% | 0 of 12 (0%) 0 to 24% |
| Title can't settle it: calls it a match | 0 of 12 (0%) 0 to 24% | 11 of 12 (92%) 65 to 99% |
| Title can't settle it: misses it | 1 of 12 (8%) 1 to 35% | 1 of 12 (8%) 1 to 35% |
Ranges are 95% Wilson intervals. Word match flags whatever the search finds, with no way to abstain.
These are first-run numbers. We set the held-out recalls aside on Nov 5, and later rule changes left these numbers the same. One held-out recall's CPSC sentence (26054) is used as an example in one rule's test.
Where it was wrong
On a wider screen, this is a table of all 38.
When a title states the model
We searched a public product index for 40 recalled models, one search each. Of 311 products found, 0 had the model in the title. So these titles are built from each recall's own words, and are not real listings. They show the matcher reads a model when one is stated, and nothing more.
| 60 recalls | Recall Reach | Word match |
|---|---|---|
| Title states the recalled model: confirmed | 52 of 60 (87%) 76 to 93% | 60 of 60 (100%) 94 to 100% |
| Model one character off: wrongly confirmed | 0 of 53 (0%) 0 to 7% | 53 of 53 (100%) 93 to 100% |
The other 8 ask instead. On 6 of them CPSC limits the recall within the model, by VIN, a code or a look-up, so the model alone can't settle it.
An earlier version confirmed 58 of these 60. That was wrong for 6 recalls CPSC limits within the model, so it asks on those now.
What Claude says about these
Claude read the 48 held-out pairs that reach a seller. Its finding is text for the seller to read. It changes nothing above.
Claude also wrote these labels, so this table shows it agrees with itself as much as with the truth. A person's audit is what would make it independent.
| Claude said | On look-alikes | On pairs the title can't settle |
|---|---|---|
| Same product | 0 of 37 (0%) 0 to 9% | 0 of 11 (0%) 0 to 26% |
| Can't tell | 22 of 37 (59%) 43 to 74% | 11 of 11 (100%) 74 to 100% |
| Different product | 15 of 37 (41%) 26 to 57% | 0 of 11 (0%) 0 to 26% |
Where it falls short: on 22 of 37 look-alikes Claude said it can't tell, so the seller still has to ask. It called none of the 48 the same product, but with Claude's own labels that isn't proof.
What this does not show
- The titles are new retail listings. Resale titles are shorter and noisier.
- This set has no true matches: every title naming a recalled line still lacked what the recall's scope needs. It can't measure finding them.
- These are small numbers, and some titles repeat across recalls. Read the ranges.
Every miss, by name
Could reach a family
- CPSC recall 26058 Werner Skagit FG 2 Piece Straight Shaft Kayak Paddle
- CPSC recall 26058 Werner Camano 2 Piece Straight Shaft Kayak Paddle
Missed, where the title can't settle it
- CPSC recall 25127 Self-Propelled Multi-Blade Mower
Reached the seller
- CPSC recall 25382 Glintex 14-Inch Kids Bicycle
- CPSC recall 25382 SUGIFT 18-Inch Kids Sport Bicycle
- CPSC recall 25382 SUGIFT 14-in Kids Bicycle
- CPSC recall 25382 HONGGE 18 Inch Kids Bicycle
- CPSC recall 25384 Montessori House-shaped Crib Bumper
- CPSC recall 25384 Padded Baby Crib Mattress
- CPSC recall 26075 Alpaca Plush
- CPSC recall 26075 Cosy Alpaca Soft Toy
- CPSC recall 26075 Alpaca Plush
- CPSC recall 26075 Alpaca Plush
- CPSC recall 26122 Youth Helmet
- CPSC recall 26122 Fox Youth Mainframe MIPS - Kids Helmet
- CPSC recall 26122 Next Youth Helmet
- CPSC recall 26122 Fox Youth Mainframe MIPS - Kids Helmet
- CPSC recall 25290 5-Mode Bluetooth Baby Swing
- CPSC recall 25220 My Bath Seat - Baby Bath Seat
- CPSC recall 25220 My Bath Seat - Baby Bath Seat
- CPSC recall 25220 Regalo® Baby Basics® Bath Seat
- CPSC recall 25220 Regalo Baby Basics Bath Seat
- CPSC recall 25287 Mini Lightning Bolt - Short Sleeve Loungewear Set
- CPSC recall 25287 Tiny Tumblers - Short Sleeve Loungewear Set
- CPSC recall 25287 Long Sleeve Tops With Pants Loungewear Pajama Sets
- CPSC recall 25123 Upsie Sling Carrier
- CPSC recall 25123 Ergobaby Upsie Sling Baby Carrier
- CPSC recall 25123 Ergobaby Upsie Sling Baby Carrier
- CPSC recall 25135 Youth Grasser Insulated Jacket
- CPSC recall 27060 Sensory Swing
- CPSC recall 25345 Momcozy Baby Lounger Bedside Sleeper
- CPSC recall 26101 Family Matching Pajama Sets
- CPSC recall 26101 Cotton Squirrel Print Children's Pajama Set
- CPSC recall 26101 Children's Pajama Set in Petite Petals
- CPSC recall 26101 Children's Pajama Set in Pink Flannel
- CPSC recall 25127 Walk Behind Self-Propelled Lawn Mower
- CPSC recall 25167 Black Diamond Recon X Avalanche Beacon
- CPSC recall 26058 Werner Skagit FG 2 Piece Straight Shaft Kayak Paddle
- CPSC recall 26058 Werner Camano 2 Piece Straight Shaft Kayak Paddle
- CPSC recall 25248 Baseus EnerGeek GX11 MiFi Portable WiFi Power Bank
Run it yourself with pnpm eval. A test fails if these numbers stop matching what the code produces. Home
Separate from these numbers: what sellers said about its flags.