torumitsutake/jev-page-verdict
User repository on GitHub
- Forks
- 0
- Last pushed
- Updated Sep 2026
Repo stats from the GitHub API, cached Oct 2026.
CommunityJev Demos & Experiments
Jev Page Verdict
Manifest V3 Chrome extension that asks nine parallel Jev questions about the page you are on to separate advertising from first-hand account, across genre, stance, publisher, product and four Noul checks.
- Stars
- 1
- Language
- JavaScript
- License
- MIT
- Published by
- Community
- Author
- torumitsutake
- Last verified
- Oct 2026
About this project
A Manifest V3 Chrome extension that answers one question about the page you are on: is this selling me something, or is it someone telling me what it was actually like? The verdict comes from Jev (TypeSafe AI), a judgement-only model: it does not generate prose, it picks, scores and answers true/false, each with a probability. That property is why this extension is small — there is no output to parse, no prompt to babysit, and a confidence number that tells it when to stay quiet.
Quoted from the project's README. Read the full README.
Works forJev for Classification
Used inJev for Software & DevTools
- Languages
- JavaScript
Highlights
- Nine questions ride one request in parallel: genre, stance, publisher, product, independence, firsthand, disclosure, thin content, machine-written.
- Affiliate links, price mentions, call-to-action phrases and schema.org types are counted in extract.js, so Jev only judges what the page is.
- At 0.75 or above a verdict is stated plainly, 0.50 to 0.75 as likely, below 0.50 Undecided; thresholds live in CONFIDENCE in questions.js.
- Google colour-coding is opt-in and three-tier: cached verdicts only, then snippets for up to 10 results, then one fetched page per result you click.
Watch out
No build step, plain ES modules loaded from the folder, and that folder must not be moved or deleted or Chrome disables the unpacked extension on its next start. The service worker background.js is also not reloaded when you press reload on the extension card. Your TypeSafe key is stored in chrome.storage.local in plain text, and the README we read points at a Security notes section that the copy we received does not contain, because it is cut off after the Google result section. Accuracy drops on non-English pages: the model's main training language is English, so instructions and criteria stay English while the state keeps the page's own language. Choice returns a distribution, so overlapping labels split probability and confidence; that is why the detailed preset's eight classes are documented as needing thresholds dropped to around 0.4, and why machine-generated is a Noul rather than a genre. Pages starting with chrome://, the Chrome Web Store, and file:// pages cannot be judged unless you allow file URLs. Snippet mode never fetches the linked page, and the fetch-on-click path gets HTML without cookies or scripts, so a body under 400 characters is rejected rather than judged. No accuracy figures, cost per page, or per-question latency are published. Copy reviewed against the repository README on 2026-10-04, not from running it.
Questions about this entry
Short answers drawn from the linked project or page.
The README gives four reasons and puts privacy first: only pages you open, plus snippets Google already showed you, are sent. Bulk fetching would send the full text of pages you never opened, on every search, which is the same position as a content script on all_urls. It also makes prompt injection practical, because search results include pages an attacker placed there, and dozens of automated requests per search look like crawling from your IP to sites you never visited.
Hedge. Choice returns a probability distribution, so two labels that overlap in meaning split the probability and confidence drops below the threshold even when the model understood the page. The README's example puts an affiliate article at 0.4 / 0.35 / 0.15 and calls that undecided. The working rule is that one question should be answerable by an expert in seconds; if a human hesitates between two labels, the labels are wrong.
The key goes in chrome.storage.local in plain text, saved from the options page, and the README sends you to a Security notes section to decide how to handle that. On the folder: Chrome keeps pointing at the path you loaded, so an unpacked extension whose folder is gone is disabled at the next start. The folder also has no build step, so you edit questions.js and press reload on the extension card.
It judges them, but with lower accuracy, and that is a documented model property rather than a bug in the extension. The official documentation states the model's main training language is English and that accuracy drops in other languages, so instructions and criteria are always English, even with the interface in Japanese, while the state stays in whatever language the page is written in. The README we read publishes no per-language accuracy numbers.
Chrome 116 or newer, which is the manifest's minimum_chrome_version, and a TypeSafe API key. There is no build step: it is plain ES modules loaded unpacked straight from the folder, so no toolchain is needed.
One request carries nine questions evaluated in parallel as speculative fan-out: genre, stance, publisher and product as Choice, independence as a Score, and firsthand, sponsored_disclosure, thin_content and machine_written as Noul. Extra questions barely cost latency, so the design adds axes instead of stacking labels onto one axis.
At 0.75 or above the verdict is stated plainly, between 0.50 and 0.75 it is stated as likely, and below 0.50 it shows Undecided and hands the call back to you. The thresholds live in CONFIDENCE in questions.js, and larger label sets thin every probability, so expect to lower them to around 0.4.
More Jev resources like this
30GitHub stars
Side-by-side game harness where Laya running locally and Jev over API answer the same typed questions for snake, kombat, runner, flappy, tetris, and human-versus-model matches.
Jev Demos & Experiments#community#jevbest#catalog-import
19GitHub stars
Live versus Tetris where Jev faces Claude Haiku 4.5, Gemini 3.8 Flash, or local open-weight Laya on the same seeded piece sequence; cleared lines arrive as garbage rows and the first to top out loses.
Jev Demos & Experiments#community#awesomejev#catalog-import
Community
trungdq88/jev-tetrisJavaScript
trungdq88/jev-tetrisJavaScript21GitHub stars
Browser top-down driving sim that sends sensor JSON to Jev every 200ms with four typed driving questions in one call, then applies confidence gating and reflex rules to steer.
Jev Demos & Experiments#community#jevbest#catalog-import
Community
vinilana/live-jevJavaScript
vinilana/live-jevJavaScript12GitHub stars
Chrome extension that labels each X post with Substance, Humor, Chit-chat, Promo or Junk plus AI-written and off-topic flags, collapsing hidden posts to one line.
Jev Demos & Experiments#community#jevbest#catalog-import
Community
bohutang/siftJavaScript
bohutang/siftJavaScript11GitHub stars
Chrome extension that labels each X post with intent plus rage-bait, synthetic, and undisclosed-ad signals from Jev, with threshold-only rendering and local caching.
Jev Demos & Experiments#community#jevbest#catalog-import
Community
manifoldor/xtagsJavaScript
manifoldor/xtagsJavaScript8GitHub stars
Chrome extension that judges Taobao and JD listings against double-weekend-off clues and stamps a PASS mark over product images it guesses come from non-weekend-off shops.
Jev Demos & Experiments#community#awesomejev#catalog-import
From the Jev guides
Original write-ups on what this entry is for.
A game tick is a state plus a legal action list. How Clash Royale, Minecraft, and the earlier Mario, Pokémon, StarCraft, and drone loops ask Jev one step at a time.
Jev Demos & Experiments#gaming#agents#control
Read article
From the Jev community
Posts from builders shipping with Jev right now.