jev.arnao.ai · model jev-1.13 via OpenRouter · private demo

Jev Sandbox

What Jev is. A model from TypeSafe that does not write text. You give it some text plus yes/no or pick-one questions, and it returns a number from 0 to 1 for each answer: how sure it is. About a quarter of a second, about two-thousandths of a cent per call. Your code, not the model, decides what to do with the number.

checking key

bank desk · synthetic data

Hundreds of cases, seconds, a penny

Pretend we are a bank's back office. Pick how many customer cases to score. Every case gets three questions. One button.

Add one question of your own (optional)

Data is synthetic, generated by bank-gen.js (the 300 set is bank-300.json). No real people, accounts, merchants or amounts. The generator plants about 5% fraud-like, 8% churn-risk, 10% urgent and a few prompt-injection cases; the rest are ordinary.

free box

Try your own

Paste any text and ask up to five yes/no questions. Nothing you type is stored.

Text to evaluate

0 / 4000

guardrails

The OpenRouter key lives only in the serverless function; the browser never sees it.
This page needs the sandbox key once in its URL (?key=). It is then kept on this device and removed from the address bar. Without it the function answers 401.
30 calls per minute per IP (a batch counts as one), a $1.00 per day spend cap per running instance (best effort; the OpenRouter key's own credit limit is the real ceiling), text capped at 4,000 characters (2,000 per batch item), 5 questions per call (4 per batch).
No input is stored or logged. Jev's probabilities are shown to you here, but in a real gate they must never be returned to the agent (score feedback doubles an adaptive attacker's reach).
Logic and arithmetic belong in code, not in a model: the September research build scored "x AND NOT x" as 0.38 satisfiable; today's build scores it 0.02. Either way, do not ask a model to do math.