Shpigford/knockoff

The Xcode project carries its own copy of the extension files; after editing the extension, run scripts/sync-safari.sh to update it before rebuilding.

How it works

Everything runs locally in a content script. No accounts, no tracking, no server round-trips on the shopping path. Each product tile's brand is resolved through this pipeline (first match wins):

Name heuristics

Unknown brands are scored on the linguistic signature of trademark-squat names: ALL-CAPS 5–9 character strings, vanishing vowel ratios, unpronounceable consonant runs, un-English letter pairs, non-Latin characters, random iNternal caPitalization. High scores are flagged, mid scores suspect. The known-brands list always vetoes the heuristics: plenty of real brands (ASICS, RYOBI, HOKA) would otherwise look suspicious. Scoring lives in src/detector.js and is deliberately readable, and tuning it is a great first contribution.

Filter levels

Actions

Filtered items can be hidden (with a floating pill showing the count and a one-click reveal), dimmed (fade + desaturate, restore on hover), or just labeled. Every badge is clickable: trust the brand, block it, show the item once, or report a misclassification.

Product detail pages get a verdict chip next to the brand byline. The page is never hidden out from under you.

Reporting misclassifications

The badge menu has one-click reporting ("this is junk" / "this is a real brand"). Reports go to a tiny Cloudflare Worker backed by D1 (report-worker/) and are reviewed by hand to improve the bundled lists. No PII: the payload is brand, verdict, ASIN, marketplace, and extension version; reporter IPs are stored only as salted hashes for rate limiting. If no endpoint is configured the extension falls back to opening a prefilled GitHub issue.

Deploying your own endpoint is four commands; see the header of report-worker/worker.js.

Contributing

The easiest, highest-value contributions are brand list fixes; see CONTRIBUTING.md. There is no build step; the extension is plain JavaScript, loadable directly from the repo.

manifest.json             MV3 manifest
data/known-brands.js      curated established brands (edit this one!)
data/chinese-major.js     established Chinese-owned brands (toggleable)
data/flagged-brands.js    seed blocklist of notorious pseudo-brands
data/generic-words.js     common title words, for unbranded detection
data/community-brands.js  bundled community allowlist snapshot (generated, don't edit)
src/detector.js           detection engine (pure logic, no DOM)
src/content.js            page scanning, badges, actions, in-page control panel
src/background.js         toolbar button → panel toggle (or options page)
options/                  settings page (rules, allow/blocklist)
report-worker/            Cloudflare Worker: reports, curation, brand-list API
safari/                   Xcode wrapper app for Safari (macOS)
site/                     knockoff.shopping landing page (static, Cloudflare)
store-assets/             Chrome Web Store images + the HTML frames that render them
scripts/                  maintenance scripts

Known limitations

• Mixed-case gibberish ("Geinxurn", "Mulwark") scores below the suspect threshold at standard level; Strict mode catches it. A bundled character bigram model would fix this properly; PRs welcome.
• Seller country-of-origin lookup (fetching seller profile addresses) is deliberately not implemented: it costs two rate-limited page fetches per product and Amazon 503s aggressive scrapers. The name-based approach needs zero network calls.
• Carousels and a few exotic tile layouts aren't scanned yet (TILE_SELECTORS in src/content.js is the extension point).
• Non-English stores are best-effort. Brand lists and the product-page chip work everywhere, but the name heuristics are English-tuned, so an unlisted local brand can slip through. Non-Latin listings (Japanese, Arabic) are skipped rather than mis-filtered, so nothing breaks on any marketplace.

Prior art

Research that shaped this design: AmazonBrandFilter (allowlist approach; its MIT-licensed community list seeded Knockoff's own), SoldBy (seller-country lookup and its rate-limit lessons), and The Markup's Amazon Brand Detector. Knockoff's contribution is combining a community allowlist with a heuristic scorer, with the allowlist as veto.

License

FSL-1.1-MIT. Code converts to MIT after two years.