Listing Risk Screen — catch brand and policy risk before you publish

A takedown costs you the listing, the inventory behind it, and often the account. This reads your listing text against your own watchlist and tells you what to fix first:

verdict	clear · review · block — an action, not a finding
reason	one label: trademark_variant, counterfeit_positioning, copyrighted_character, prohibited_claim, restricted_category, unsubstantiated_claim, compatibility_reference, thin_listing, no_signal
near_matches	which watchlisted names were found in the text, and where to look
signals	the per-signal probabilities behind the verdict, so you can re-threshold without paying again

Point it at any dataset — an Amazon or Shopify scraper, a supplier feed, your own CSV of drafts — paste in your watchlist, and run it.

The whole point: the spellings a keyword search cannot find

A CONTAINS "Nike" rule is free and you already have it. It also finds nothing when the listing says something else. On a 260-listing corpus where every brand was disguised, a plain substring search over the watchlist scored:

accuracy 0.835   recall 0.660   precision 0.849

That is the bar. Every one of these was read correctly as the brand, and none of them contains the brand as text:

in the listing	read as
Niike	Nike
A1rP0ds	AirPods
J-B-L	JBL
Dys on	Dyson
Dyosn	Dyson
Aplpe	Apple
IuIuI3mon	Lululemon

And the opposite mistake, which costs you just as much

Flagging everything is not screening. A watchlisted word used as ordinary English is not a brand use, and none of these was blocked:

in the listing	verdict
"apple-shaped silicone charm, moulded to look like the fruit"	clear
"designed in-house by our founder, Stanley Okafor"	clear
"named after the physicist Freeman Dyson"	clear
"you will marvel at how little room this takes in a bag"	review
"tough enough to take down to the barbie for an afternoon"	review

The last two are the honest limit: a word that is both a brand and ordinary English gets held for you to glance at rather than cleared outright. Eight of the sixty clean listings in the test corpus land in review for exactly this reason. None of them is blocked, and none needs a rewrite — it costs you a look.

It also separates a legitimate compatibility reference — "a case for AirPods Pro, not made by Apple" — from a listing claiming to be the brand. The first goes to review so you can check your wording once; only the second is blocked.

What was measured

260 labelled listings, thirteen adversarial case families:

Verdict accuracy, three buckets	0.931
Infringing listings wrongly cleared	0
Clean listings wrongly blocked	0
Unreadable listings wrongly cleared	0
Clean listings sent to review needlessly	13 of 60

Nine of the thirteen families are perfect: exact brand use, counterfeit positioning, copyrighted characters, prohibited claims, restricted categories, thin listings and multi-signal combinations all 20/20.

The mistakes it does make go one step in the cautious direction — into review, where you look at it. That is deliberate: a needless review costs two minutes and a missed infringement costs the account.

What it does not do

This screens text. It is a risk signal, not a legal opinion and not a clearance. Design-patent and image rights live in the product photo and are not screened at all — a listing marked clear here has not been cleared of those. Every output row carries that disclaimer, because a report that reads as legal advice is worse than no report.

Two limits worth knowing before you buy:

• Three-letter brands are hard. A one-character swap on a short name is surfaced as a near match, so the listing goes to review, but it may not be read as the brand itself.
• A variant written in lowercase is not surfaced as a near match. Exact matches are found in any casing.

Billing

You are charged per listing screened — never per listing returned, and never for one it failed to judge. Those come back in the dataset marked billed: false with the error on the row, so nothing vanishes silently.

Before any work starts, the run is capped by your own charge limit, so it cannot bill its way into a wall halfway through and hand you a half-screened catalogue. If more than a fifth of the calls fail, the run fails rather than returning a report whose gaps read as clean.

Input

screenPolicy is required and is the whole product:

{
  "own_brands": ["Your Brand"],
  "watchlist": {
    "brands": ["Apple", "AirPods", "Anker", "Nike", "Dyson"],
    "franchises": ["Disney", "Mickey Mouse", "Pokemon"],
    "prohibited_terms": ["FDA approved", "cures", "100% safe", "best seller"]
  },
  "restricted_categories": ["lithium batteries", "supplements", "cosmetics"]
}

own_brands are yours and are never flagged. A name that is not on the list is a name nobody looks for, so put your real competitors and the franchises your category attracts on it.