Grounds

A small Chrome extension (Manifest V3) that reads the article you are on and shows a badge saying how well argued it is: whether the page backs its claims with reasons and structure, or just asserts them. It uses the Jev API (TypeSafe's "System One" model, which returns typed decisions instead of prose).

Grounds is not a fact checker. It does not judge whether the conclusions are true, only how the article argues. An opinion column can be well argued; a confident sounding rant is not.

The badge floats in the bottom right corner (inside a shadow DOM, so page styles cannot break it). It shows one label plus Jev's confidence, coloured like a traffic light:

1. Extraction. Readability (the engine behind Firefox and Safari Reader View) isolates the main article body and drops nav, sidebars, comments and related boxes. It runs on a clone of the DOM, so the cost is negligible.
2. Local gate. If Readability returns a long, coherent body (about 1,500 characters or more) the page counts as an article. If there is too little text (under about 500 characters) Grounds stays quiet and makes no Jev call.
3. Safety check for ambiguous pages. When extraction is uncertain, the single Jev call also asks a choice question (single article, feed or listing, or other). The badge shows only if Jev confirms a single article.
4. Verdict. One score call to Jev on a four level argumentation rubric. The raw score maps to the badge label and is shown with Jev's confidence.

One article means one Jev call (about 100 ms). There are no per sentence requests.

1. Get a Jev API key from TypeSafe.
2. Clone this repository (or download the ZIP).
3. Open chrome://extensions and turn on Developer mode (top right).
4. Click Load unpacked and select this folder.
5. Open the popup, paste your Jev API key, then either press Analyze page on an article, or turn on Auto analyze to run it on articles as you read.

There is no build step. It is plain HTML, CSS and JavaScript.

There is no API key in this repository, and none ships with the extension. Each user provides their own Jev key. It is stored only in your browser (chrome.storage.local), it never syncs, and it is used only by the background service worker. The web page never sees it.

To score an article, Grounds sends the extracted article text to the Jev API (https://api.typesafe.ai) over HTTPS, together with your key. In auto mode this happens for every article you read. Grounds:

• sends only the extracted article text, not the full page HTML or form fields;
• makes the request from the background service worker, so your key is never exposed to the page;
• can make privileged cross origin requests only to api.typesafe.ai (see host_permissions);
• does not store, log, or send your data anywhere else.

Anything you send is handled under TypeSafe's own terms. Do not run it on pages you consider private or confidential unless you are comfortable with that.

The parts you are likely to change live in a few constants:

• Rubric and prompt: background.js, ARGUMENT_INSTRUCTIONS and ARGUMENT_CRITERIA. These define what "well argued" means and tell the model to ignore rhetoric.
• Endpoint and model: background.js, JEV_ENDPOINT and JEV_MODEL.
• Thresholds: content.js, STRONG_MIN, MIN_CHARS, MAX_CHARS.
• Labels and colours: content.js, LEVELS.

• How well argued is still a judgment. Treat the badge as a signal, not a verdict. That is why Jev's confidence is shown.
• Telling genuine structure apart from persuasive rhetoric is hard, so borderline articles can move between levels 2 and 3.
• Soft paywalls that show only a teaser are judged on the visible portion.
• It does not run on special pages (chrome://…, the Web Store, PDFs). That is a Chrome limitation.

• Article extraction by Mozilla Readability (Apache 2.0), vendored in Readability.js.
• Scoring by the Jev model (TypeSafe). This project is not affiliated with or endorsed by TypeSafe.
• This extension's own code is released under the MIT License. See LICENSE.