zod-jev

How we reviewed this: we ran its 78 tests in Docker on Node 24 with no network and read the request builder. We made no Jev calls.

What it does

zod-jev, by jomatsu, extends Zod 4 with semantic rules:

const z = createJevZod();
const Ticket = z.semantic(z.object({ subject: z.string(), body: z.string() }), [ /* rules */ ]);

Shape rules stay in Zod. Meaning rules — “does this body contain personal data?”, “is this price plausible for this item?” — become Nouls, and a single parseAsync sends every rule of that schema in one request, then converts probabilities that miss your threshold into ordinary Zod issues.

That last part is the good design decision: callers keep using safeParse, error.issues and everything else they already have, rather than learning a new failure shape. The default messages are localised, with the instructions sent to the model kept in English deliberately.

All 78 tests pass offline.

What to know

The state is the object being validated, so whatever your schema covers is what leaves the process. For a support ticket that’s fine; for a user record, decide first.

And a probability isn’t a verdict: you pick the threshold, and the library is explicit that this is your call.

Verdict

A clean, small idea executed without ceremony — one request per parse, familiar error objects, no new vocabulary. If your codebase already runs on Zod, this is the least intrusive way to add a judgment to a boundary you already validate.

For the same instinct in Ruby, see ruby_llm-typesafe.

See how it compares with other tools in Best Jev tools, tested hands-on.

Review updated Sep 20, 2026. Numbers quoted from the project are its author's own; we don't publish our own measurements of Jev.