What's new in 3.0

This is a rebuild rather than an update. The way results are presented has changed, and so has what the extension does on the network.

Every finding now says how strong its own evidence is. Five levels, from something that was fetched and checked, down to something the model merely remembers. Previously a judgement based on the model's recollection of a publisher looked exactly like a measurement. It no longer does, and that distinction is the point of this release.

There is no single reliability score. Results come as four separate readings — what the text does, who published it, what independent sources say, and whether the page's own links work. A reading that could not run says so and why, rather than showing zero. "We could not look" and "we looked and it was bad" are different answers.

Two new checks, both off until you turn them on, each explained before you do:
- Link checking follows the links a page already contains to see whether they resolve.
- Lateral reading asks your chosen AI provider what independent sources say about the publisher, and cites them.

Nothing loads from anyone else any more. Earlier versions pulled styles and fonts from third-party servers, so opening Settings contacted them; one earlier version also routed requests through a third-party proxy. All of that is gone. Your API key stays on the machine you typed it on and is never included in browser sync. Page text is not stored.

Crabe assesses text only — never images, video or audio, including whether they are authentic or shown in their original context. Every result says so.

And every result carries one more line: Crabe supports critical judgement, it does not replace it. Any model can be confidently wrong. Read the page critically, and read the assessment critically too.

Also: bring your own key for Google Gemini or Anthropic Claude, or run a local model through Ollama so the page text never leaves your machine. The panel now survives clicking elsewhere, so a long analysis is no longer lost.