JevLabDECISION EXPERIMENTS
01 / EXPERIMENT

Give Jev a judgment call.

How this works ↗
Checking sign-in…

The input

Text in. Typed decisions out.
START WITH AN EXAMPLE

Choice picks an option. Score rates ordered levels. Noul estimates the probability of “yes”. Question syntax ↗

Model & input format
One direct call · 20s deadline

The judgment

No run selected
?

What will Jev decide?

Run an example or edit the evidence and questions. The actual answer, probabilities and confidence will appear here.

No simulated answers. No hidden model fallback.

Run notebook

Last 100 runs · private to your account

Sign in to load saved runs.

02 / READING THE RESULTS

A decision is evidence to inspect.

Each run sends the state and every question directly to TypeSafe. Questions are evaluated independently against the same state; one answer cannot feed another within the call.

Probability ≠ confidence

Choice shows a distribution over your options. Score uses levels starting at zero and may land between them. Noul returns a yes probability. Choice and Score also return TypeSafe’s separate confidence estimate.

Structured does not mean correct

Test ambiguous and missing-evidence cases. Keep an “unknown” option where it makes sense. A selected answer does not authorize actions or prove its own accuracy.

Runs survive reloads

The server saves inputs, timestamps and terminal results. A stopped worker fails visibly; a 20-second deadline ends stalled runs. Reopening a run reads its saved record and never repeats a paid call.

Connection & cost

Live evaluations require a server-side TypeSafe API key. Credentials are never included in the saved request.

Cost is an estimate from reported input tokens at $0.042 per million, with output free. It is not a provider invoice.

TypeSafe model documentation ↗