SlopBlock

SlopBlock is a local-first Chrome/Chromium extension for finding secondhand gems on Facebook Marketplace by removing everything that isn't one: dropshipped listings, Amazon/Temu-style retail slop, bait prices, sponsored ads, external redirects, store and dealer ads, service/job/MLM spam, scam scripts, counterfeit goods, repeated listing floods, and dead sold/wanted posts — so real secondhand gems are easier to find.

No account, backend, analytics, monetization, or remote rule updates. Filtering happens in the browser against visible Marketplace text; only your settings are stored locally with chrome.storage.local.

How it decides

Explainable weighted rules with a three-band action ladder — see docs/FILTERING.md for the full model:

• Hide is reserved for near-certain signals (sold/ISO posts, scam scripts, sponsored ads, confirmed bait pricing, replica language, repeat floods).
• Dim (fogged but visible) covers medium evidence — nothing is removed on a hunch.
• Label annotates mild suspicion, and every filtered card can show its reasons.

Gem-hunting defaults: anything mentioning IKEA is hidden outright (flat-pack isn't a gem — one tap switches to Filter/Off), and brand-new/in-box titles dim by default (retail flips). Both act instantly from card text with no network wait, and both are fully reversible tri-state quick filters.

Repeated-listing floods are collapsed, not punished: the first copy always stays visible and only repeats beyond it hide ("repeat of a visible listing, 3 of 9 identical"). Same title alone never counts — search results full of "iPhone 12 128GB" from different sellers at different prices are untouched.

Deep scan catches description-only spam right in the feed. Cards only show price/title/location, so a listing whose only tell is a Wayfair catalog link in its description looks innocent while browsing. Deep scan background-fetches the listing page from facebook.com (the same request as clicking it, throttled and viewport-first), reads the seller-written fields from the page data, and applies the verdict to the card before you ever open it. On by default; one switch to turn off; nothing leaves the browser.

Photos are analyzed on-device — no OCR models, no uploads. Retailer catalog photos are product-on-pure-white (Amazon requires it), so a canvas check catches Amazon/Wayfair picture copy-pastes locally in microseconds. A perceptual hash catches repost floods that reuse the same photo while rotating price and location — a pattern no text rule can see. And a "Find photo online" button on flagged cards opens Google Lens for a full reverse image search, strictly when you click it.

Item detail pages get their own scan. Feed cards only show price/title/location, so listings whose evidence lives in the description (retailer catalog links, order language, scam scripts) are scored when you open them: outbound links are decoded from Facebook's l.facebook.com wrappers, truncated "See more" descriptions are expanded, and a verdict banner with reasons and one-click Allow/Disable appears at the top of the listing — the page itself is never hidden. Feed-time deep scan runs up to 10 fetches in parallel over Facebook's HTTP/2 connection (rate-limited, viewport-first, prunes what you scroll past, backs off if Facebook pushes back), so a page is vetted in seconds.

Tested against real listings

The rules are tuned and gated against real-world text, not intuition:

• eval/corpus/ holds 195 labeled entries: real listings fetched from public marketplace pages (furniture, electronics, cars, free stuff, tools…), real dealer/scam/dropship/counterfeit text collected from consumer-protection reports and community documentation, plus curated edge cases (IKEA resales, "cash or venmo", "$15 each", phone numbers, "no dealer fees here"…).
• npm run eval scores the whole corpus at every strength in both card view (price/title/location — what feed cards actually show) and detail view; it reports false positives, misses, and per-rule noise.
• npm run eval:gate (part of npm run verify) fails the build if any legit listing gets dimmed or hidden at balanced strength. Current state: 0 false positives at any strength, 98% of slop actioned in detail view.
• npm run e2e loads the built extension into real Chromium against a high-fidelity Marketplace DOM fixture (structure reconstructed from public scraper sources) and runs 56 end-to-end checks: hiding, flood collapse, sponsored-cell removal, item detail-page banners (including the See-more/link-wrapper handling), infinite scroll, on-card badges, allow-item persistence, popup summary and settings sync, options rule tester.

Using it

• The popup shows what happened on the current page: scanned/hidden counts, top triggers with one-click Disable, hidden examples with one-click Allow item, strength/action controls, and quick filters. Vendor filters are three-way — Off / Filter / Hide all. IKEA ships in Hide all (gem-hunting default); Amazon/Temu sources, brand-new items, and liquidation ship in Filter. Everything saves instantly.
• A small pill in the corner of Marketplace pages shows how many listings were filtered; click it for counts and a "Show hidden" preview where every hidden card displays its reason with Allow/Disable buttons.
• The options page has every built-in rule with its exact pattern, weight, and confidence (searchable, individually toggleable), a rule tester for pasting listing text, custom block/vendor/allow terms, and JSON export/import. All changes autosave.

Development

npm install
npm test          # typecheck + 104 unit tests
npm run eval      # score the real-listing corpus (add -- --verbose for details)
npm run bench     # per-scan hot-path performance budget (scoring + DOM extraction)
npm run e2e       # build + end-to-end tests in real Chromium (needs npx playwright install chromium once)
npm run build     # production build into dist/
npm run verify    # test + eval gate + build + audit + package checks
npm run package   # verify + write release/slopblock.zip
npm run diagnostics -- path/to/diagnostics.json
node e2e/screenshots.mjs   # UI screenshots (light+dark) into .tmp-tests/screens

Load in Chrome

1. npm run build
2. Open chrome://extensions, enable Developer mode.
3. "Load unpacked" → select dist/.
4. Open Facebook Marketplace.

Tuning false positives

Fast to deep: on-card badge buttons → popup "This page" (disable a trigger / allow an item) → popup quick filters → strength/action mode → options rule groups → individual rules → custom allow terms → Copy diagnostics + npm run diagnostics for offline analysis. Details in docs/FILTERING.md.

Privacy

SlopBlock sends nothing anywhere. No backend, telemetry, ads, payments, or accounts. Copied diagnostics strip URL tracking params and redact phone/email patterns, and only leave the browser if you paste them somewhere yourself. See PRIVACY.md.