Question 1

What stops it from scoring my whole timeline before the first label appears?

Accepted Answer

The extension scores posts as they enter or approach the viewport and does it in batches, so labels are usually ready before you stop on a tweet. Scores are cached by URL in the worker's KV for 90 days, a daily cron lists and deletes expired keys so the namespace cannot grow without bound, and stale hits are dropped on read.

Question 2

Is this just a Cloudflare Worker I could deploy there myself?

Accepted Answer

Same shape: a standard Workers-compatible project written with Wrangler, and the local loop is wrangler dev. Production here means that same worker running on open-compute, the author's self-hosted platform compatible with the Workers model. Endpoints are /v1/score for one post, /v1/score/batch capped at 50 items, /health, and OPTIONS for CORS from chrome-extension:// and localhost.

Question 3

How accurate are the AI-written and paraphrase calls?

Accepted Answer

The README publishes no precision, recall or false-positive rate for any of the three judgments, so treat the labels as a filter rather than a measurement. What it documents is the design: judgment combines the author's handle, display name and bio with the post text, plus the parent post for replies, with the three-band cut at 0.65 and 0.35.

Question 4

Where do my API key and my browsing data actually go?

Accepted Answer

The TypeSafe key lives as a worker secret set with wrangler secret put TYPESAFE_API_KEY and never in the extension, which only POSTs extracted posts to your worker. An optional second API_KEY gates client access. The popup's language setting, blur or hide filter, separate post-versus-reply thresholds and local scan counts are stored only in the browser.

Question 5

How do I run it from source?

Accepted Answer

bun install, then bun run dev for both packages, bun run dev:extension for the WXT extension alone, or bun run dev:worker for wrangler on http://127.0.0.1:8788. bun run test runs the worker vitest suite. Packaging is bun run extension:zip, which writes packages/extension/.output/tweet-911-chrome.zip. For local secrets, copy .dev.vars.example to .dev.vars.

Question 6

What does the scoring worker expose?

Accepted Answer

GET /health returning {"ok": true}, POST /v1/score for a single post, POST /v1/score/batch taking { "items": [...] } capped at 50 items, and OPTIONS on any path for CORS from chrome-extension:// and localhost. url is required because it is the cache key; send text and/or image URLs, plus author and in_reply_to when you have them. The model defaults to jev-latest.

Question 7

Can I change the thresholds, and where are those settings kept?

Accepted Answer

Yes. Scores are 0 to 1 and the README's own cut-offs are fixed constants at 0.65 and 0.35, giving likely_ai, uncertain or likely_human, with the same pattern for the solicitation and paraphrase judgments. The popup is where you set your own: separate post and reply thresholds, a blur or hide filter, language, and local scan counts, all of which stay in the browser.

Tweet 911

About this project

Highlights

Questions about this entry

More Jev resources like this

From the Jev guides

From the Jev community