Quality and accuracy
How every check is tested, and what the numbers mean. Each number below was measured on October 22, 2026, on that day’s copy of the official lists, and carries that date. Where something has not been measured, the page says so.
At a glance
Measured October 22, 2026 on that day’s copy of each list. “Caught” means flagged as a match or for review; “accuracy on clean companies” is the share of invented companies, clean by construction, that were correctly cleared.
Results come from fixed rules, not from AI
- Every verdict is computed by ordinary code from what the list searches returned. The same supplier and the same lists give the same answer every time.
- AI writes one thing: the plain-language paragraph explaining a finding. The code that sets a verdict takes only the check results as its input, so nothing the AI writes can move a verdict.
- On a customs notice, AI only reads the words on the page. The hold type and the response deadline are then worked out by code, from the fields read.
- Every finding cites its source and names the lists searched, so it can be checked by hand.
Where the data comes from, and how fresh it is
Each check matches against a local copy of the official list, refreshed daily from the government source (the UFLPA Entity List via OpenSanctions’ structured copy of the DHS publication). Every result records which copy it used. If a copy grows older than its cutoff, the check stops and says so, rather than answering from stale data. The dates below are read live:
How the name matching is measured
Matching a typed name to a listed company can go wrong in two directions: missing a listed company typed slightly differently, or flagging a clean one. Both are measured together, because a matcher can look perfect on one by giving up the other.
Catching listed companies. Real names from each list were typed in six changed forms. The forced-labor lists are the UFLPA Entity List and CBP Withhold Release Orders; the sanctions lists are OFAC’s lists and the U.S. Consolidated Screening List (companies only: people, ships and aircraft on those lists are outside a supplier screen).
| How the name was changed | Forced-labor lists | Sanctions lists |
|---|---|---|
| Letters swapped (a typo) | 100% (61 of 61) | 100% (99 of 99) |
| Legal form written differently | 100% (56 of 56) | 100% (3 of 3) |
| Legal form left off | 100% (56 of 56) | 100% (3 of 3) |
| Punctuation removed | 100% (58 of 58) | 100% (9 of 9) |
| Name split into two words (Ningbo → Ning bo) | 100% (18 of 18) | 100% (27 of 27) |
| A place name added (… Shenzhen) | 100% (63 of 63) | 100% (101 of 101) |
Clearing clean companies. 900 invented company names, clean by construction and spread across 64 legal forms from around the world, were screened. All 900 were correctly cleared on the forced-labor lists, and 846 of 900 (94%) on the sanctions lists. The rest share one uncommon industry word with a listed company, such as “Toys” or “Extrusions”; they are flagged for a person to review, never reported as a match. Tightening the matcher to clear them was tested and rejected: it would miss real listed companies typed with one extra word.
A name resemblance is never presented as proof. The government lists publish a name, aliases and a country, so the report says how close the match is and what it would take to confirm it.
How the CPSC certificate mapping is audited
CPSC’s flagged list maps HS codes to the safety rules that may apply and the citation codes CPSC accepts; the certificate check reads it for a code that comes with the product (from a product page or a report) and reaches every one of those rules from the product questions alone. That chain was walked end to end on October 22, 2026, for every code on CPSC’s current flagged list, not a sample:
- All 589 flagged HS codes, across 23 product categories, map to a category and a suggested safety rule, except 11 imitation-jewelry codes, which are left without a suggestion on purpose (overlapping lead-content rules with no single clean citation) and are shown that way.
- 0 citation codes that CPSC does not publish can be produced.
- All 79 safety rules offered had their scope read in the regulation itself (the Code of Federal Regulations), not only in CPSC’s summaries.
CPSC says its flagged list does not cover every code that may need a certificate, and that a code on it is only likely to include a regulated product. So a product can need a certificate without a flagged code, and a flagged code does not guarantee a product is regulated. The results say so.
What is tested on every change
- More than 2,000 automated tests run on every proposed change, including the verdict logic, the certificate questionnaire walked from an empty form, and the customs-notice routing.
- The CPSC Registry file is checked against CPSC’s published column schema, and that check is itself tested by planting known errors and confirming each one is caught.
What the checks cannot do
- They search the official lists named on each result, not every list in the world.
- They cannot confirm a company’s identity beyond public registries. A company that cannot be corroborated is reported as “could not corroborate”, never as fake.
- They are research and screening tools, not legal or customs advice, and not a licensed customs broker. A broker or a customs attorney makes the filing and the final call.
Checking a report is genuine
Anyone who receives a report can confirm it was issued by this service, and see its date and outcome line, at /verify, using the reference ID and the six-character code printed on it (reports issued since September 21, 2026). How information is handled in each tool is set out on How client data is handled.