Question 1

Does it rewrite my text when it dislikes it?

Accepted Answer

No. It evaluates supplied writing against named checks and shows a plain-language verdict, a labeled answer confidence and a relevant sentence. It does not generate a rewrite, and it does not establish that a passage is factually correct. The author frames it as an independent community project, not an official TypeSafe product.

Question 2

Why separate questions instead of one overall writing score?

Accepted Answer

Each enabled check becomes a separate typed question and they are all submitted in one primary batch, so you get a named verdict with confidence and an evidence sentence rather than one opaque number. The interface deliberately keeps counts in the filter controls and has no overall writing score anywhere.

Question 3

Will the public demo spend my credits, or does it leak my key?

Accepted Answer

The deployment carries no server key on purpose, because a public URL configured with one would let any visitor spend its credits, so the demo serves deterministic simulated results labeled as demo mode. For real verdicts you paste your own key, kept in that browser's localStorage and sent only as a header to the app's own /api/judge endpoint.

Question 4

What happens when the model is unsure or returns something malformed?

Accepted Answer

Those cases are separated from a real verdict. Uncertain answers are labeled Pass or Issue at low confidence, and missing or malformed answers say Needs review. Changing the text, the checks or the mode marks old verdicts stale and names what changed, so simulated results cannot pass as live ones.

Question 5

Can I add a check of my own, and where is it stored?

Accepted Answer

Yes. Add a custom check with a name, a question and an answer type, either Yes/No or a fixed list of options, then declare which answers count as a problem and optionally describe what a good result looks like. The form previews the exact question, options and issue criteria before saving. Custom checks live in that browser's localStorage, editing one keeps its id, and saving a changed check makes previous results stale.

Question 6

Which model and wire format does the app actually call?

Accepted Answer

Jev answers structured questions rather than writing a critique. The server adapter speaks TypeSafe's state, model and questions wire format against jev-latest, in lib/jevClient.ts, using noul for Yes/No checks and choice for named options. Live requests go through the app's own /api/judge endpoint to TypeSafe, never directly from the browser to the provider, and the topbar reports the last run's latency and token use.

Question 7

How accessible is the interface, and how far does that claim go?

Accepted Answer

Automated axe-core checks run against every page and state with WCAG 2.2 AA rules plus best practices and report no violations, and the README calls that a floor rather than a guarantee. Underneath it aims at keyboard and screen-reader use: a skip link as the first tab stop, a command palette on Cmd or Ctrl K, probability tracks named as meters, verdicts carrying an icon and a word as well as colour, 4.5:1 text contrast and 44px minimum touch targets.

Clarity Judge — BunsDev

About this project

Highlights

Questions about this entry

Other listings of this project

More Jev resources like this

From the Jev guides

From the Jev community