Question 1

Fourteen labelled samples is a thin basis for two thresholds. How much should I trust them?

Accepted Answer

Treat them as a starting point, which is what the author says. The reported ranges are 0.87 to 0.94 is_slop for engagement bait, 0.10 to 0.13 for genuine posts, 0.23 to 0.27 for corporate PR and 0.26 to 0.35 for the ambiguous group, with is_corporate separating corporate PR at 0.98 from bait at 0.09. The claim of zero false positives means nothing genuine scored above 0.35, on fourteen samples. The thresholds are fitted to the exact question wording in server/jev.js, so change the wording and re-run ./test/run.sh before trusting them again.

Question 2

Why does this need a local server instead of being just an extension?

Accepted Answer

Because anything bundled into a Chrome extension is readable by everyone who installs it. The key therefore lives in the server directory and never reaches the browser, and the requests are made from the local proxy on 127.0.0.1:8787. That is also the stated reason it is not on the Chrome Web Store: shipping there means either leaking a key or making every user paste their own. The trade-off is that the proxy must be running, and when it is not the extension degrades to showing everything.

Question 3

Will this break on the next LinkedIn build?

Accepted Answer

Probably, at some point, and the README says so. Every class name LinkedIn ships is hashed and rotates each build. The only durable hook is an attribute selector matching the main feed component key, so when stamps stop appearing that selector is the first thing to check. It is a content script that observes posts as they scroll into view and stamps them in place, which means it depends on LinkedIn's DOM as much as on the model.

Question 4

Does it ever hide or suppress a post?

Accepted Answer

No. The post stays readable underneath the stamp, and the failure behaviour is explicit: if the proxy is down or Jev errors, every post stays visible, because a broken judgement must never hide a real post. The stated limits are narrower ones, such as text-only judgement and English-only thresholds. You can preview the stamp animation without LinkedIn at all by serving the test directory and opening the standalone preview page.

Question 5

How do I install and run it?

Accepted Answer

Node 20 or newer, Chrome, and a TypeSafe key. Clone the repo, copy .env.example to .env with the key in it, then cd server and npm start to bring the proxy up on 127.0.0.1:8787. In Chrome turn on developer mode, load the extension folder unpacked, and open linkedin.com/feed. The popup reports how many posts were judged and whether the proxy is connected.

Question 6

Can I try the stamp without LinkedIn open?

Accepted Answer

Yes. Serve the test directory with python3 -m http.server 8899 and open /test/stamp-preview.html. The animation runs 380 ms, entering at 5.2x scale with motion blur, squashing to 0.84, overshooting to 1.06 and settling at a random tilt, with a shock ring on the bottom-out and a recoil 150 ms later. It uses transform and opacity only and honours prefers-reduced-motion.

Question 7

What is actually in the repository?

Accepted Answer

Three directories. extension/ is the Chrome MV3 extension with the content script, popup, stamp CSS and icons. server/ is the local proxy that holds the key, runs the local keyword rules and calls Jev. test/ holds the labelled samples, the threshold runner and the standalone stamp preview. The project is MIT licensed.

Jev Linkedin Slop Filter

About this project

Highlights

Questions about this entry

More Jev resources like this

From the Jev guides

From the Jev community