X · Xiaohongshu · Weibo
X · 小红书 · 微博
You can already spot the obvious ads.
This one is for the ads you can't spot.
明着打的广告你自己就能看出来。
这个插件专治看不出来的那种。
Ad — openly selling. Native — a pitch dressed as personal experience. Funnel — free material traded for your contact. NSFW — explicit content and the accounts farming it (can be switched off on its own). High-confidence hits collapse themselves; hover a label to see the full probability split.
硬广明着卖货 · 软广伪装成个人体验的推广 · 引流用免费资料换你的联系方式 · 黄推色情内容与色情引流(可单独关闭)。 高置信度的会自动折叠,鼠标悬停能看到完整概率分布和判断依据。
A blocklist kills all of these too:
关键词会把这些一起误杀:
Take the essence, discard the dross — same goes for people.
Blocklist hits 「精华」(serum / essence) → flagged as skincare · Jev: not an ad, 0.07
取其精华去其糟粕,做人不也是这个道理
关键词「精华」命中 → 误判为护肤广告 · Jev:非广告 0.07
The Professor has a pile of legendary skins and none of them look like him
Blocklist hits a game item whose name contains the word · Jev: not an ad, 0.03
第五人格超话 教授一堆高级皮除了精华紫都没有自己的脸
关键词「精华」命中 → 误判 · Jev:非广告 0.03(游戏道具名)
A worthwhile investigation into the economics of the adult film industry and the labour protections in it
Any blocklist kills this · Jev: not explicit, 0.07 — it is reporting
日本AV产业的经济规模与从业者权益保障,一篇值得读的调查报道
关键词屏蔽必杀 · Jev:非色情内容 0.07(新闻报道)
And it still catches a native ad with no trigger word in it at all:
同时它认得出一个敏感词都没有的软广:
These 4 are what actually calmed my sensitive skin down — safe even right after a laser session
No price, no link, no coupon · Jev: native ad, 0.73
这4瓶是我敏感肌用下来觉得好用的退红抗炎维稳精华液,光电术后也能用
无价格、无链接、无优惠券 · Jev:软广 0.73
343 posts scraped from live feeds, every label reviewed by hand.
343 条真实抓取的中文帖子,人工逐条复核标注。
| Metric | Result |
|---|---|
| Ad vs not-an-ad | 94% |
| Recall — ads that got flagged | 96% |
| False positives on ordinary posts | 7% |
| Four-way accuracy | 89% |
| 指标 | 结果 |
|---|---|
| 广告 / 非广告 判对率 | 94% |
| 广告召回(该标出的标出了) | 96% |
| 误报(正常帖被标) | 7% |
| 四类细分准确率 | 89% |
Three sensitivity levels let you pick your own trade-off between missing ads and over-flagging. Click any label to report a mistake — those reports are what tunes it.
灵敏度分三档,可以自己在「宁可漏过」和「宁可错杀」之间选。 判错了点一下角标就能上报,数据会用来改进判断。
Paid plans open soon. Checkout is closed while our payment provider finishes reviewing the account. In the meantime the extension is free to use with a launch key — grab one from the launch thread, or email jevadguide@gmail.com.
付费套餐即将开放。支付服务商还在审核账号,结账暂时关闭。 在此期间可以用体验码免费使用 —— 从发布贴里领一个,或发信到 jevadguide@gmail.com。
$1.99 / month/ 月
$19 / year/ 年
Both plans renew automatically — monthly or yearly — until you cancel. Cancel any time from the customer portal linked in your receipt email; you keep access until the period you have already paid for ends, and are not charged again. Refunds revoke the licence key. Full terms.
Card, PayPal, Apple Pay and Google Pay, in your local currency. Your licence key is emailed the moment payment clears. Only posts you actually scroll past get classified, and a post that is already popular is judged once for everyone.
两个套餐都会自动续费(按月或按年),直到你取消为止。可随时从收据邮件里的客户门户取消; 取消后已付费的当期仍可正常使用,不会再次扣款。退款会同时撤销授权码。完整条款见服务条款。
支持信用卡、PayPal、Apple Pay、Google Pay,按你的本地货币结算。付款后授权码会立即发到你的邮箱。 按次计费,刷过去没看的帖子不消耗额度,热门内容全网只判一次。
Chrome, Edge or any Chromium browser. Download the extension and:
1. Open chrome://extensions and turn on Developer mode
2. Unzip, then click "Load unpacked" and pick the folder
3. Open the extension's Details → Extension options and paste your licence key
A Chrome Web Store listing is in review. Until it is approved, the download above is the only way to install — it is the same package.
Chrome / Edge 等 Chromium 内核浏览器。下载 安装包 后:
1. 打开 chrome://extensions → 右上角开启「开发者模式」
2. 解压 zip,点「加载已解压的扩展程序」选中该文件夹
3. 点扩展的「详细信息」→「扩展程序选项」,填入授权码
Chrome 应用商店版本正在审核中。上线之前,上面的下载是唯一的安装方式,内容完全相同。
No. Only the text of posts that actually enter your viewport is sent for a verdict, and it
is discarded once the verdict comes back — the post itself is never stored. No account details,
no record of which URLs you visit, no cookies. The extension does not request
tabs, history or cookies, and does not use <all_urls>.
See the privacy policy.
不会。只有进入你可视区域的帖子的文字内容会被发送用于判断,判完即丢,不保存原文。
不读取账号信息、不记录访问过哪些网址、不碰 Cookie。扩展没有申请
tabs、history、cookies 权限,也没有使用 <all_urls>。
详见隐私政策。
Open the customer portal link in the receipt email Creem sent you and cancel there — it takes one click and needs no email to us. Your licence keeps working until the period you already paid for runs out, and you are not charged again. If you would rather we did it, email jevadguide@gmail.com from the address you bought with.
Yes. On the measured set, 7% of ordinary posts pick up a label and 11% of the four-way calls are wrong. That is why only high-confidence hits collapse — everything else just gets a small label and stays readable. Click it to report a mistake; it takes a second.
打开 Creem 发给你的收据邮件里的客户门户链接,在那里点一下就能取消,不需要联系我们。 授权码在你已付费的当期内继续有效,之后不会再扣款。想让我们代办的话, 用购买时的邮箱发信到 jevadguide@gmail.com。
会。实测 7% 的正常帖会被标注,四类细分有 11% 判错。所以默认只折叠高置信度的, 其余只贴角标不影响阅读。判错了点角标上报,一秒的事。
At 300 posts a day the monthly 20,000 lasts you over two months. Popular posts share one verdict across everyone, so they are not billed again.
每天刷 300 条的话,月卡 20,000 次够用两个月还有余。 热门内容在所有用户之间共享判定结果,不重复计费。
Not yet. Mobile browsers do not run extensions, and inside the apps there is no hook at all.
暂不支持。移动端浏览器不支持扩展,App 内更没有办法。
Its threshold is far more conservative than the ad one. Sex education, relationship talk and reporting about the adult industry all stay unflagged. It can be switched off on its own.
阈值比广告识别保守得多。性教育、情感话题、关于色情产业的新闻报道都不会被标。 这个功能可以在设置里单独关闭。