DEV Community

zacfire
zacfire

Posted on

TypeSafe Jev vs “just use JSON schema on an LLM”

A common confusion: “Can’t I get structured output from GPT/Claude with JSON schema? Then what is Jev for?”

What JSON-schema prompting still is

You ask a text generator to emit tokens that happen to look like JSON (or that a constrained decoder forces into a schema). You still:

  • generate language tokens
  • parse / validate afterward
  • hope the semantic choice is right (schema-valid ≠ correct decision)

Useful. Not the same product shape as TypeSafe’s System One pitch.

What Jev is aiming at

TypeSafe describes Jev as: software state + typed questions (Choice / Score / Noul-style) → calibrated probabilities, composed in your code. It’s framed as a decision engine, not a chat LLM that you prompt into JSON.

So:

LLM + JSON schema Jev (System One, vendor shape)
Core loop generate tokens → parse typed questions → probabilities
Role general language model constrained decision-only model
Good for drafts, tools, flexible text fast structured judgments in-app

Don’t collapse them into “Jev is just JSON mode.”

Personal (non-official) landing

Crawlable explainer + use cases: https://aitier.app/jev/

Personal tier poster chip: jev on aitier.app Models — not an official leaderboard.

Top comments (0)