Jev's typed decisions are replacing LLM-as-judge in eval pipelines
TypeSafe's System One model returns schema-valid answers in ~400ms at 1/76th an LLM's per-call cost — and just became Vercel's fastest-adopted launch.
Copy markdown
A model that decides, instead of writing
System One models skip token-by-token generation — a parallel sampler returns your whole answer in one query, as a schema-defined Choice, Score, or boolean. TypeSafe reports a 0% structured-output and tool-call error rate, so the JSON-extract-validate-retry glue around LLM classifiers just disappears.
~0.4s a case, output tokens free
End-to-end latency runs 70–500ms; TypeSafe clocks about 0.4s per workflow case against ~10s for GPT-5.6 Terra and ~38s for Claude Opus 5. Input is $0.042/MTok and output is unmetered — roughly $0.0004 a call, about 1/76th of GPT-5.6 Terra, the line between scoring every request and not.
Vercel's fastest-adopted model ever
Vercel wired Jev into AI Gateway within three days of launch and says it reached ~13% of paid teams in its first 24 hours — more than double any prior model. You call it with one key, with automatic fallbacks and request traces, exactly like any other model.
Build this weekend: swap your grader for Jev
A wave of eval tooling wired Jev in as the judge this week — Openlayer's jevals ("agent evals and guardrails in one request"), jev-as-a-judge, and pull requests into ChainForge and Pipecat. Point your existing eval set at a Jev cascade for calibrated typed verdicts at a fraction of the LLM-judge bill and latency.
The catch: closed, gated, decisions-only
Jev is proprietary and API-only (POST api.typesafe.ai/v1/systemone), still behind an early-access waitlist, and it returns only Choice/Score/boolean answers — a decision layer, not a generator, so you keep an LLM for anything that writes text. TypeSafe says to calibrate it to your schema rather than trust it cold.
Elsewhere: two similar judges aren't two judges
A widely-shared analysis today argues two similar AI graders reportedly fail together about 7.7x more often than independence predicts — so bolting on a second correlated judge, LLM or Jev, buys far less safety than the math implies. If you double up graders, make them different.