The model behind the sifter
What is TypeSafe Jev?
Jev is a System One decision model from TypeSafe AI. "System One" means it answers typed questions fast, with a structured verdict, instead of generating prose. It is the engine that scores every row jev-curate pours through it.
Pricing: 0.042 USD per million input tokens, output unmetered. Endpoint https://api.typesafe.ai/v1/systemone.
{ "state": "let f(x) = x^2 ...",
"questions": [
"Does this chain of thought entail the answer?",
"Rate reasoning depth." ] }
answers:
noul 0.05 (circular reasoning)
score 4.8 (reasoning depth, conf 0.95)
The three answer types
Jev does not chat. It returns one of three typed shapes, and the caller decides what to build on top: thresholds, routes, or gates.
| Type | Shape | What it is for |
|---|---|---|
| Choice | one option + confidence | routing, classification, gating to a heavier eval |
| Score | 1 to 5 + confidence | ratings: quality, depth, relevance |
| Noul | calibrated probability | odds of a property: circular logic, dishonesty, breakage |
Every preset in jev-curate is a bundle of such questions. The reasoning-math preset asks whether the chain of thought entails the answer and how deep the reasoning is. The anti-sycophancy preset asks rated honesty questions. The code-correctness preset asks whether the program is sound. None of them ask for an essay, because no preset needs one.
Built for batch, priced for batch
The model profile makes the batch economics clear. Input costs 0.042 USD per million tokens, output is unmetered, and throughput sits near 250k tokens per second with 1,200 requests per minute. TypeSafe's own workflow benchmarks claim 193.6x faster and 444.6x cheaper than chat-style judging on System One tasks. That is why jev-curate can afford to evaluate every row instead of a sample: the verdict is nearly free.
The client respects those limits with a token bucket. Measured locally against a mock server, one node sustains 24.0 rows per second; the benchmark page explains the number and the arithmetic behind scaling it to the 1,500 rows/sec cluster target.
If you came from a forum post asking "what is jev ai", the honest one-liner is: TypeSafe AI's model for typed decisions, and this site's tool is one consumer of it. The glossary defines every term, and the use case pages show the workflows.