Jev Repl

Run npx jev-repl, or install it globally with npm i -g jev-repl and then run jev. With TYPESAFE_API_KEY set you get live answers; without it the REPL starts in mock mode and answers are simulated locally.

Three: noul returns the probability of yes on a 0-1 scale, choice returns the selected label with per-label probabilities and confidence, and score returns a probability-weighted level with its legend, per-level probabilities, and confidence.

No. Without TYPESAFE_API_KEY it starts in mock mode, where answers are simulated locally and deterministically rather than predictively. The :key command switches the session to live calls.

The :cost command counts tokens per question and on both sides of the wire. Rates are supplied by you, for example :cost 0.20/1.00 or JEV_PRICE=0.20/1.00, given as dollars per million tokens, input then output. Without rates it counts tokens and stops there.

Yes. A page saved with :save can be run by the jev subcommands: jev run, jev eval, jev json, jev cost, jev ts, jev rust, and jev check. Each reads a page or stdin, prints one thing, and reports success through its exit status.

It re-asks every question after each turn, from the first turn onward, and draws one line per question so you can see when a rubric noticed rather than only where it ended up. The last column names every turn the answer changed.

Tokens are estimated from the request body at roughly four characters per token, so they describe a shape rather than an invoice. A live answer carries counted usage, and the REPL prices that instead of the estimate.