Question 1

Is this just a lint rule set with an LLM behind it?

Accepted Answer

The seven built-in checks, covering hedging, em dash overuse, clarity up front, filler phrases, tone consistency, passive voice overuse and actionability, are Jev questions, each enabled check becoming a separate typed question submitted in one primary batch, with noul for yes/no and choice for named options. It does not rewrite your text or establish that a passage is factually correct.

Question 2

Will my draft get sent through someone else's TypeSafe account?

Accepted Answer

Only through your own key. The public demo deliberately runs with no server key, so a visitor who has not supplied one sees deterministic simulated results labeled as demo mode. A key you paste yourself stays in that browser's localStorage and travels only as a request header to the deployment's own judge endpoint; the deployment never displays it again or stores it.

Question 3

What is the number on each card, a quality score or a probability?

Accepted Answer

Neither an overall grade nor a single score; the compact overview deliberately keeps no overall writing score. Each card carries a plain-language verdict, a labeled confidence, and a supporting sentence, with raw yes probability or option probabilities plus the exact question behind Details. Uncertain answers read Pass with low confidence or Issue with low confidence, and malformed ones read Needs review.

Question 4

Should I trust the hosted demo's results or run it locally?

Accepted Answer

The demo's verdicts are simulated, not measured: the README says they show how the interface behaves, not Jev's performance, and switching between demo and live marks earlier verdicts stale so simulated ones never pass as live. For real verdicts you need Node 20 or newer plus pnpm, since the install guard permits pnpm only and you should not create a second lockfile, then set TYPESAFE_API_KEY in .env.local.

Question 5

What are the seven built-in checks?

Accepted Answer

Hedging language, em dash usage, clarity up front, filler phrases, tone consistency, passive voice overuse and actionability. They are editable writing preferences rather than universal rules, since a hedge can express real uncertainty and passive voice can be appropriate.

Question 6

Can I add my own check?

Accepted Answer

Yes. Add a custom check with a name, question and answer type of Yes/No or a fixed option list, then declare which answers count as a problem and optionally describe a good result. Custom checks live in that browser's localStorage, can be edited without replacing their ID, and undo restores the most recently removed check.

Question 7

What does the model actually return?

Accepted Answer

Jev answers structured questions instead of writing a critique: noul for Yes/No and choice for named options. The server adapter uses TypeSafe's state, model and questions wire format with jev-latest, and the default 70% flagging threshold can be moved to re-filter cards locally without another model call.

Clarity Judge — TypeSafeAI

About this project

Highlights

Quickstart

Questions about this entry

From the Jev guides

From the Jev AI model community